Text embedding and reranking are foundational to modern information retrieval systems, powering applications such as semantic search, recommendation systems, and retrieval-augmented generation (RAG). However, current approaches often face key challenges—particularly in achieving both high multilingual fidelity and task adaptability without relying on proprietary APIs. Existing models frequently fall short in scenarios requiring nuanced semantic understanding across multiple languages or domain-specific tasks like code retrieval and instruction following. Moreover, most open-source models either lack scale or flexibility, while commercial APIs remain costly and closed. Qwen3-Embedding and Qwen3-Reranker: A New Standard for Open-Source Embedding Alibaba’s Qwen Team has unveiled the Qwen3-Embedding and Qwen3-Reranker Series—models that set a new benchmark in multilingual text embedding and relevance ranking. Built on the Qwen3 foundation models, the series includes variants in 0.6B, 4B, and 8B parameter sizes and supports a wide range of languages (119 in total), making it one of the most versatile and performant open-source offerings to date. These models are now open-sourced under the Apache 2.0 license on Hugging Face, GitHub, and ModelScope, and are also accessible via Alibaba Cloud APIs. These models are optimized for use cases such as semantic retrieval, classification, RAG, sentiment analysis, and code search—providing a strong alternative to existing solutions like Gemini Embedding and OpenAI’s embedding APIs. Technical Architecture Qwen3-Embedding models adopt a dense transformer-based architecture with causal attention, producing embeddings by extracting the hidden state corresponding to the [EOS] token. Instruction-awareness is a key feature: input queries are formatted as {instruction} {query}<|endoftext|>, enabling task-conditioned embeddings. The reranker models are trained with a binary classification format, judging document-query relevance in an instruction-guided manner using a token likelihood-based scoring function. The models are trained using a robust multi-stage training pipeline: Large-scale weak supervision: 150M synthetic training pairs generated using Qwen3-32B, covering retrieval, classification, STS, and bitext mining across languages and tasks. Supervised fine-tuning: 12M high-quality data pairs are selected using cosine similarity (>0.7), fine-tuning performance in downstream applications. Model merging: Spherical linear interpolation (SLERP) of multiple fine-tuned checkpoints ensures robustness and generalization. This synthetic data generation pipeline enables control over data quality, language diversity, task difficulty, and more—resulting in a high degree of coverage and relevance in low-resource settings. Performance Benchmarks and Insights The Qwen3-Embedding and Qwen3-Reranker series demonstrate strong empirical performance across several multilingual benchmarks. On MMTEB (216 tasks across 250+ languages), Qwen3-Embedding-8B achieves a mean task score of 70.58, surpassing Gemini and GTE-Qwen2 series. On MTEB (English v2): Qwen3-Embedding-8B reaches 75.22, outperforming other open models including NV-Embed-v2 and GritLM-7B. On MTEB-Code: Qwen3-Embedding-8B leads with 80.68, excelling in applications like code retrieval and Stack Overflow QA. For reranking: Qwen3-Reranker-0.6B already outperforms Jina and BGE rerankers. Qwen3-Reranker-8B achieves 81.22 on MTEB-Code and 72.94 on MMTEB-R, marking state-of-the-art performance. Ablation studies confirm the necessity of each training stage. Removing synthetic pretraining or model merging led to significant performance drops (up to 6 points on MMTEB), emphasizing their contributions. Conclusion Alibaba’s Qwen3-Embedding and Qwen3-Reranker Series present a robust, open, and scalable solution to multilingual and instruction-aware semantic representation. With strong empirical results across MTEB, MMTEB, and MTEB-Code, these models bridge the gap between proprietary APIs and open-source accessibility. Their thoughtful training design—leveraging high-quality synthetic data, instruction-tuning, and model merging—positions them as ideal candidates for enterprise applications in search, retrieval, and RAG pipelines. By open-sourcing these models, the Qwen team not only pushes the boundaries of language understanding but also empowers the broader community to innovate on top of a solid foundation. Check out the Paper, Technical details, Qwen3-Embedding and Qwen3-Reranker. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 95k+ ML SubReddit and Subscribe to our Newsletter. Asif RazzaqWebsite |  + postsBioAsif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.Asif Razzaqhttps://www.marktechpost.com/author/6flvq/A Step-by-Step Coding Guide to Building an Iterative AI Workflow Agent Using LangGraph and GeminiAsif Razzaqhttps://www.marktechpost.com/author/6flvq/From Clicking to Reasoning: WebChoreArena Benchmark Challenges Agents with Memory-Heavy and Multi-Page TasksAsif Razzaqhttps://www.marktechpost.com/author/6flvq/Mistral AI Introduces Mistral Code: A Customizable AI Coding Assistant for Enterprise WorkflowsAsif Razzaqhttps://www.marktechpost.com/author/6flvq/NVIDIA AI Releases Llama Nemotron Nano VL: A Compact Vision-Language Model Optimized for Document Understanding

In my last Fast Company column, I shared my reasons for manufacturing my electric trucks in the U.S. I’m not alone. While near-shoring to North America has been underway for several years, the current tariff shifts and shipping complexities make U.S. manufacturing a higher priority still. However, there were 292,825 factories in the U.S. as of 2021. Of those, 846 employ 1,000 people or more. Some of these are my engineering firm’s clients, giving us a front row seat to the complexity of retrofitting an existing factory to full U.S. manufacturing. While building a new factory is expensive and lengthy, these companies’ tasks are more difficult still. There are good reasons for making the shift as quickly as possible. Moving to most or fully U.S. manufacturing brings higher visibility, faster response time, and higher resilience to supply chain disruption, as well as greater protection from tariff shifts and geopolitical change. But if you’re early in the process, here’s my advice for your transition: Determine a priority ranking for the refining and raw materials you shift to in-country and North American sourcing. Give highest ranking to categories including defense, high value items (such as steel, aluminum, and rare minerals, etc.), and consumer safety items (such as pharmaceutical components, etc.). Be more strategic in the offshore suppliers you continue using for non-advanced manufacturing by prioritizing closer and more geographic-friendly locations for production and shipment such as Mexico and Argentina. Utilize government-backed capital, where possible, for extracting/mining minerals and metals such as lithium, red mud, magnesium, etc. Beyond the high-ranking product categories, move to domestic suppliers for primary materials such as steel, aluminum, cement, and plastics. Likewise, reduce offshoring of technical staff as well as raw materials, where possible. Use all means possible to become power independent through solar production, micro-grids, and nuclear power production. Consider creating a 4-year completion bonus for military vets. Hire vets wherever possible, as they make great workers and entrepreneurs. Likewise, we can press for future policy changes that best support Made in America manufacturing, as follows: Encourage ship building in the U.S., as well as creating new means of automated freight transit. Work towards transformation plans for government-funded R&D to include more attractive loans, rebates, and grants, as well as programs for tax-free status for intellectual property during commercialization, to incent and support organizations making the shift. Consider energy rebates to U.S. manufacturers and distributors to make American manufacturing more cost-effective and viable.   Create policies to include the cost of offshore staff in tariff calculations. Expand trade relationships with Caribbean nations for products such as sugar, avocados, bananas, etc. Avoid or even ban foreign ownership of the food supply chain. Create fair competition for government contracting. Make health supplements and homeopathic medicines tax deductible, to promote a healthy workforce. While it may not be readily evident, these policy changes are related to successful reshoring. In all, we need larger scale, lower costs, and more automated and simplified mechanisms for product manufacturing. These issues, in my experience, are as equally important as the raw materials we require. We need increased support for niche manufacturing. In my opinion, we also need deregulation, and increased access to land (particularly in the west; the federal government owns great quantities of the available land, which is choking available supply). I believe we need better education, self-reliance, health, and incentive structures to get the capital, entrepreneurs, and workers for Made in America manufacturing. Who’s with me? Matthew Chang is the founding partner of Chang Robotics.

The ways roads are used, with ever larger and heavier vehicles, have dramatic consequences on the environment – and electric cars are not the answer Today, there is an average of 37 tonnes of road per inhabitant of the planet. The weight of the road network alone accounts for a third of all construction worldwide, and has grown exponentially in the 20th century. There is 10 times more bitumen, in mass, than there are living animals. Yet growth in the mass of roads does not automatically correspond to population growth, or translate into increased length of road networks. In wealthier countries, the number of metres of road per inhabitant has actually fallen over the last century. In the United States, for instance, between 1905 and 2015 the length of the network increased by a factor of 1.75 and the population by a factor of 3.8, compared with 21 for the mass of roads. Roads have become wider and, above all, much thicker. To understand the evolution of these parameters, and their environmental impact, it is helpful to trace the different stages in the life of the motorway.  Until the early 20th century, roads were used for various modes of transport, including horses, bicycles, pedestrians and trams; as a result of the construction of railways, road traffic even declined in some European countries in the 19th century. The main novelty brought by the motorway was that they would be reserved for motorised traffic. In several languages, the word itself – autostrada, autobahn, autoroute or motorway – speaks of this exclusivity.  Roman roads varied from simple corduroy roads, made by placing logs perpendicular to the direction of the road over a low or swampy area, to paved roads, as this engraving from Jean Rondelet’s 19th‑century Traité Théorique et Pratique de l’Art de Bâtir shows. Using deep roadbeds of tamped rubble as an underlying layer to ensure that they kept dry, major roads were often stone-paved, metalled, cambered for drainage and flanked by footpaths, bridleways and drainage ditches Like any major piece of infrastructure, motorways became the subject of ideological discourse, long before any shovel hit the ground; politicians underlined their role in the service of the nation, how they would contribute to progress, development, the economy, modernity and even civilisation. The inauguration ceremony for the construction of the first autostrada took place in March 1923, presided over by Italy’s prime minister Benito Mussolini. The second major motorway programme was announced by the Nazi government in 1933, with a national network planned to be around 7,000 kilometres long. In his 2017 book Driving Modernity: Technology, Experts, Politics, and Fascist Motorways, 1922–1943, historian Massimo Moraglio shows how both programmes were used as propaganda tools by the regimes, most notably at the international road congresses in Milan in 1926 and Munich in 1934. In the European postwar era, the notion of the ‘civilising’ effect of roads persevered. In 1962, Valéry Giscard d’Estaing, then‑secretary of state for finances and later president of France, argued that expanded motorways would bring ‘progress, activity and life’. This discourse soon butted up against the realities of how motorways affected individuals and communities. In his 2011 book Fighting Traffic: The Dawn of the Motor Age in the American City, Peter D Norton explores the history of resistance to the imposition of motorised traffic in North American cities. Until the 1920s, there was a perception that cars were dangerous newcomers, and that other street and road uses – especially walking – were more legitimate. Cars were associated with speed and danger; restrictions on motorists, especially speed limits, were routine.  Built between 1962 and 1970, the Westway was London’s first urban motorway, elevated above the city to use less land. Construction workers are seen stressing the longitudinal soffit cables inside the box section of the deck units to achieve the bearing capacity necessary to carry the weight of traffic Credit: Heritage Image Partnership Ltd / Alamy To gain domination over cities, motor vehicles had to win priority over other street uses. Rather than restricting the flow of vehicles to minimise the risk of road accidents, a specific infrastructure was dedicated to them: both inner‑city roads and motorways. Cutting through the landscape, the motorway had, by definition, to be inaccessible by any other means of transport than motorised vehicle. To guarantee the fluidity of traffic, the construction of imposing bridges, tunnels and interchanges is necessary, particularly at junctions with other roads, railways or canals. This prioritisation of one type of user inevitably impacts journeys for others; as space is fragmented, short journeys are lengthened for those trying to navigate space by foot or bicycle.  Enabling cars to drive at around 110–140km/h on motorways, as modern motorways do, directly impacts their design, with major environmental effects: the gradient has to be gentle (4 per cent), the curves long (1.5km in radius) and the lanes wide, to allow vehicles to overtake each other safely. As much terrain around the world is not naturally suited to these requirements, the earthworks are considerable: in France, the construction of a metre of highway requires moving some 100m3 of earth, and when the soil is soft, full of clay or peat, it is made firmer with hydraulic lime and cement before the highway’s first sub‑layers are laid. This material cost reinforces the criticisms levelled in the 1960s, by the likes of Jane Jacobs and Lewis Mumford, at urban planning that prioritised the personal motor vehicle. When roads are widened to accommodate more traffic, buildings are sliced and demolished, as happened in Dhaka’s Bhasantek Road in 2021 Credit: Dhaka Tribune Once built, the motorway is never inert. Motorway projects today generally anticipate future expansion (from 2×2 to 2×3 to 2×4 lanes), and include a large median strip of 12m between the lanes, with a view to adding new ones. Increases in speed and vehicle sizes have also translated into wider lanes, from 2.5m in 1945 to 3.5m today. The average contemporary motorway footprint is therefore 100 square metres per linear metre. Indeed, although the construction of a road is supposed to reduce congestion, it also generates new traffic and, therefore, new congestion. This is the principle of ‘induced traffic’: the provision of extra road capacity results in a greater volume of traffic. The Katy Freeway in Texas famously illustrates this dynamic. Built as a regular six‑lane highway in the 1960s, it was called the second worst bottleneck in the nation by 2004, wasting 25 million hours a year of commuter time. In 2011, the state of Texas invested US$2.8 billion to fix this problem, widening the road to a staggering total of 26 lanes. By 2014, the morning and afternoon traffic had both increased again. The vicious circle based on the induced traffic has been empirically demonstrated in most countries: traffic has continued to increase and congestion remains unresolved, leading to ever-increasing emissions. In the EU, transport is the only sector where greenhouse gas emissions have increased in the past three decades, rising 33.5 per cent between 1990 and 2019. Transport accounts for around a fifth of global CO₂ emissions today, with three quarters of this figure linked to road transport. Houston’s Katy Freeway is one of the world’s widest motorways, with 26 lanes. Its last expansion, in 2008, was initially hailed as a success, but within five years, peak travel times were longer than before the expansion – a direct illustration of the principle of induced traffic Credit: Smiley N Pool / Houston Chronicle / Getty Like other large transport infrastructures such as ports and airports, motorways are designed for the largest and heaviest vehicles. Engineers, road administrations and politicians have known since the 1950s that one truck represents millions of cars: the impact of a vehicle on the roadway is exponential to its weight – an online ‘road damage calculator’ allows you to compare the damage done by different types of vehicles to the road. Over the years, heavier and heavier trucks have been authorised to operate on roads: from 8‑tonne trucks in 1945 to 44 tonnes nowadays. The European Parliament adopted a revised directive on 12 March 2024 authorising mega‑trucks to travel on European roads; they can measure up to 25 metres and weigh up to 60 tonnes, compared with the previous limits of 18.75 metres and 44 tonnes. This is a political and economic choice with considerable material effects: thickness, rigidity of sub‑bases and consolidation of soil and subsoil with lime and cement. Altogether, motorways are 10 times thicker than large roads from the late 19th century. In France, it takes an average of 30 tonnes of sand and aggregate to build one linear metre of motorway, 100 times more than cement and bitumen.  The material history of road networks is a history of quarrying and environmental damage. The traces of roads can also be seen in rivers emptied of their sediment, the notches of quarries in the hills and the furrows of dredgers extracting sand from the seabed. This material extraction, arguably the most significant in human history, has dramatic ecological consequences for rivers, groundwater tables, the rise of sea levels and saltwater in farmlands, as well as biodiversity. As sand is ubiquitous and very cheap, the history of roads is also the history of a local extractivism and environmental conflicts around the world.  Shoving and rutting is the bulging and rippling of the pavement surface. Once built, roads require extensive maintenance – the heavier the vehicles, the quicker the damage. From pothole repair to the full resurfacing of a road, maintenance contributes to keeping road users safe Credit: Yakov Oskanov / Alamy Once roads are built and extended, they need to be maintained to support the circulation of lorries and, by extension, commodities. This stage is becoming increasingly important as rail freight, which used to be important in countries such as France and the UK, is declining, accounting for no more than 10 per cent of the transport of commodities. Engineers might judge that a motorway is destined to last 20 years or so, but this prognosis will be significantly reduced with heavy traffic. The same applies to the thousands of motorway bridges: in the UK, nearly half of the 9,000 highway bridges are in poor condition; in France, 7 per cent of the 12,000 bridges are in danger of collapsing, as did Genoa’s Morandi bridge in 2018. If only light vehicles drove on it, this infrastructure would last much longer. This puts into perspective governments’ insistence on ‘greening’ the transport sector by targeting CO2 emissions alone, typically by promoting the use of electric vehicles (EVs). Public policies prioritising EVs do nothing to change the mass of roads or the issue of their maintenance – even if lorries were to run on clean air, massive quarrying would still be necessary. A similar argument plays out with regard to canals and ports, which have been constantly widened and deepened for decades to accommodate ever-larger oil tankers or container ships. The simple operation of these infrastructures, dimensioned for the circulation of commodities and not humans, requires permanent dredging of large volumes. The environmental problem of large transport infrastructure goes beyond the type of energy used: it is, at its root, free and globalised trade. ‘The material life cycle of motorways is relentless: constructing, maintaining, widening, thickening, repairing’ As both a material and ideological object, the motorway fixes certain political choices in the landscape. Millions of kilometres of road continue to be asphalted, widened and thickened around the world to favour cars and lorries. In France, more than 80 per cent of today’s sand and aggregate extraction is used for civil engineering works – the rest goes to buildings. Even if no more buildings, roads or other infrastructures were to be built, phenomenal quantities of sand and aggregates would still need to be extracted in order to maintain existing road networks. The material life cycle of motorways is relentless: constructing, maintaining, widening, thickening, repairing, adding new structures such as wildlife crossings, more maintaining.  Rising traffic levels are always deemed positive by governments for a country’s economy and development. As Christopher Wells shows in his 2014 book Car Country: An Environmental History, car use becomes necessary in an environment where everything has been planned for the car, from the location of public services and supermarkets to residential and office areas. Similarly, when an entire economy is based on globalised trade and just‑in‑time logistics (to the point that many service economies could not produce their own personal protective equipment in the midst of a pandemic), the lorry and the container ship become vital.  The final stage in the life of a piece of motorway infrastructure is dismantling. Like the other stages, this one is not a natural outcome but the fruit of political choices – which should be democratic – regarding how we wish to use existing roads. Dismantling, which is essential if we are to put an end to the global extractivism of sand and aggregates, does not mean destruction: if bicycles and pedestrians were to use them instead, maintenance would be minimal. This final stage requires a paradigm shift away from the eternal adaptation to increasing traffic. Replacing cars and lorries with public transport and rail freight would be a first step. But above all, a different political and spatial organisation of economic activities is necessary, and ultimately, an end to globalised, just-in-time trade and logistics. In 1978, a row of cars parked at a shopping centre in Connecticut was buried under a thick layer of gooey asphalt. The Ghost Parking Lot, one of the first projects by James Wines’ practice SITE, became a playground for skateboarders until it was removed in 2003. Images of this lumpy landscape serve as allegories of the damage caused by reliance on the automobile Credit: Project by SITE Lead image: Some road damage is beyond repair, as when a landslide caused a large chunk of the Gothenburg–Oslo motorway to collapse in 2023. Such dramatic events remind us of both the fragility of these seemingly robust infrastructures, and the damage that extensive construction does to the planet. Credit: Hanna Brunlöf Windell / TT / Shutterstock 2025-06-03 Reuben J Brown Share

It may be time to suck it up — and we don't just mean the carbon in the atmosphere. No, we're talking about reckoning with the possibility that our attempts at capturing the greenhouse gas to stave off climate disaster are already hopelessly behind schedule, New Scientist reports, if they're not in vain entirely.To illustrate, here're some simple numbers. The CO2 removal industry expects to hit a milestone of removing one million metric tons of CO2 this year. And companies across the globe have bought carbon credits to remove 27 million more, according to data from CDR.fyi cited in the reporting (more on these carbon credit schemes in a moment).That sounds like a lot, but it really isn't. As New Scientist notes, the Intergovernmental Panel on Climate Change — the leading authority on these issues — concluded in a 2022 report that we need to be removing up to 16 billion tons of carbon, not millions, each year to keep the rise in global temperature from exceeding 1.5 degrees Celsius (2.7 degrees Fahrenheit) of warming by the middle of the century, past which the most drastic effects of climate change are believed to be irreversible."It's not scaling up as fast as it would need to if we are going to reach multiple gigatons by 2050," Robert Höglund at Marginal Carbon, a climate consultancy based in Sweden, told the magazine. Carbon capture is not the be-all and end-all. The fact remains that humanity needs to drastically reduce its emissions, which probably means reorganizing society — or at least its energy production and consumption — as we know it. Simply removing the CO2 that's already there is more like a band-aid that buys us a little time; eventually, we'll need to rip it off.For these reasons, some critics fear that carbon capture — and even more drastic interventions, like attempting to dim the Sun — could distract from the climate change's systemic causes. But there's a lot of enthusiasm for the approach all the same, both from scientists and investors. The IPCC acknowledged in its 2022 report that carbon removal was "unavoidable" — as in, essential to meeting climate targets.One popular method of carbon removal is called direct air capture, which involves sucking the carbon straight from the air using massive industrial facilities. A more circuitous approach that's gaining steam involves extracting CO2 out of the ocean, freeing up room for the world's largest carbon sink to passively absorb even more of the greenhouse gas. All of these initiatives, though, are basically just getting off the ground. And the corporate investment, which once promised billions of dollars in cash, seems to be cooling. More than 90 percent of all carbon removal credits sold this year were bought by a single company, Microsoft, New Scientist notes, probably to gloss over its egregious energy bill it's accrued from building loads of AI datacenters.This also touches on the fact that the practice of buying carbon credits can be used as a means of corporate greenwashing. By paying to another firm to "certify" that they will remove a certain amount of carbon at some undetermined point in the future, a company can report a greener carbon balance sheet without actually reducing its emissions.In any case, staking the industry's hopes on corporate munificence is a dicey prospect indeed."I have been raising the alarm for about a year and a half," Eli Mitchell-Larson at Carbon Gap, a UK carbon dioxide removal advocacy organisation, told New Scientist. "If we're just waiting for the waves of free philanthropic money from corporations to fill a hole on their sustainability report, we're not really going to solve the problem."More on climate change: Scientists Just Found Who's Causing Global WarmingShare This Article

Microsoft and Google are taking distinctive approaches with AI agents in their productivity suites, and enterprises need to account for the differences when formulating digital labor strategies, analysts said. In recent months, both companies have announced a dizzying array of new agents aimed at extracting value from corporate documents and maximizing efficiency. The tech giants have dropped numerous hints about where they’re headed with AI agents in their respective office suites, Microsoft 365 and Google Workspace. Microsoft is reshaping its Copilot assistant as a series of tools to create, tap into, and act on insights at individual and organizational levels. The Microsoft 365 roadmap lists hundreds of specialized AI tools under development to automate work for functions such as HR and accounting. The company is also developing smaller AI models to carry out specific functions. Google is going the opposite way, with its large-language model Gemini at the heart of Workspace. Google offers tools that include Gems for workers to create simple custom agents that automate tasks such as customer service, and Agentspace in Google Cloud to build more complex custom agents for collaboration and workflow management. At the recent Google I/O developer conference, the company added real-time speech translation to Google Meet. “For both, the goal is to bring usable and practical productivity and efficiency capabilities to work tools,” said Liz Miller, vice president and principal analyst at Constellation Research. But the differing AI agent strategies are heavily rooted in each company’s philosophical approaches to productivity. Although Microsoft has long encouraged customers to move from its traditional “perpetual-license” Office suite to the Microsoft 365 subscription-based model, M365 notably retains the familiar desktop apps. Google Workspace, on the other hand, has always been cloud-based. Microsoft users are typically a bit more tethered to traditional enterprise work styles, while Google has always been the “cloud-first darling for smaller organizations that still crave real-time collaboration,” Miller said. When it comes to the generative AI models being integrated into the two office suites, “Google’s Gemini models are beating out the models being deployed by Microsoft,” Miller said. “But as Microsoft expands its model ‘inventory’ in use across M365, this could change.” Microsoft has an advantage, as many desktop users live in Outlook or Word. The intelligence Copilot can bring from CRM software is readily available, while that integration is more complex in the cloud-native Google Workspace. “Microsoft still has an edge in a foundational understanding of work and the capacity to extend Copilot connections across applications as expansive as the Office suite through to Dynamics, giving AI a greater opportunity to be present in the spaces and presentation layers where workers enjoy working,” Miller said. Microsoft’s Copilot Agents and Google’s Gems and Agentspace are in their early stages, but there have been positive developments, said J.P. Gownder, a vice president and principal analyst on Forrester’s Future of Work team. Microsoft recently adopted Google’s A2A protocol, which makes it easier for users of both productivity suites to collaborate and unlock value from stagnant data sitting on other platforms. “That should be a win for interoperability,” Gownder said. But most companies that are Microsoft shops have years or decades of digital assets that hold them back from considering Google, he said. For example, Excel macros, pivot tables, and customizations cannot be easily or automatically migrated to Google Sheets, he said. “As early as this market is, I don’t think it’s fair to rank either player — Microsoft or Google — as being the leader; both of them are constructing new ecosystems to support the growth of agentic AI,” Gownder said. Most Microsoft Office users have moved to M365, but AI is helping Google is making inroads into larger organizations, especially among enterprises that are newer and less oriented toward legacy Microsoft products, said Jack Gold, principal analyst at J. Gold Associates. Technologies like A2A blur the line between on-premises and cloud productivity. As a result, “Google Workspace is no longer perceived as inferior, as it had been in the past,” Gold said. And for budget-constrained enterprises, the value of AI agent features is not the only consideration. “There is also the cost equation at work here, as Google seems to have a much more transparent cost structure than Microsoft with all of its user classes and discounts,” Gold said. Microsoft does not include Copilot in its M365 subscriptions, which vary in price depending on the type of customer. The Copilot business subscriptions range from $30 per user per month for M365 Copilot to $200 per month for 25,000 messages for Copilot Studio, which is also available under a pay-as-you-go model. Google has flat subscription pricing for Workspace, starting at $14 per user per month for business plans with Gemini included.

While Europe races to phase out fossil fuels and electrify everything from cars to heating systems, it’s turning a blind eye to a reliable and proven source of clean energy lying right beneath our feet.  Geothermal energy offers exactly what the continent needs most: clean, local, always-on power. Yet, it only accounted for 0.2% of power generation on the continent in 2024. Something needs to change. The recent blackout in Spain, triggered by a failure in the high-voltage grid, serves as a warning shot. While solar and wind are vital pillars of decarbonisation, they’re variable by nature. Without steady, around-the-clock sources of electricity, Europe risks swapping one form of energy insecurity for another. A much bigger wake-up call came in 2022 when Russia launched a full-scale invasion of Ukraine. For years, European governments had built an energy system dependent on imports of natural gas. When that stack of cards shattered, it triggered an energy crisis that exposed the vulnerable underbelly of Europe’s power system.  The answer to these problems lies, in part, a few kilometres underground. According to the International Energy Agency, geothermal energy has the potential to power the planet 150 times over. But it’s not just about electricity — geothermal can also deliver clean, reliable heat. That makes it especially valuable in Europe, where millions of homes already rely on radiators and district heating systems, many of them still powered by natural gas. Geothermal plants also come with a smaller footprint. They require far less land than an equivalent solar farm or wind park. What’s more, the materials and infrastructure needed to build them — like drilling rigs and turbines — can be largely sourced locally. That’s a sharp contrast to solar panels and batteries, most of which are imported from China.   Geothermal energy is not theoretical. It doesn’t require scientific breakthroughs. We’ve been drilling wells and extracting energy from the Earth for centuries. The know-how exists, and so does the workforce. Decades of oil and gas exploration have built a deep bench of geologists, drillers, reservoir engineers, and project managers. Instead of letting this expertise fade, we can redeploy it to build geothermal plants. The infrastructure, such as drilling rigs, can also be repurposed for a cleaner cause. Geothermal could be the ultimate redemption arc for oil and gas. Sure, drilling deep isn’t cheap — yet. But a new crop of startups is rewriting the playbook. Armed with everything from plasma pulse drills to giant radiators, these companies could finally crack the cost barrier — and make geothermal available pretty much anywhere. Just as SpaceX disrupted a sclerotic rocket industry with its cheap launches, these startups are poised to succeed where the geothermal industry has failed.  All that’s missing is investment. While billions are being funnelled into high-risk technologies like fusion or nuclear fission reactors, funding for geothermal tech is minuscule in comparison, especially in Europe. Yet, unlike those technologies, geothermal is ready right now.   If Europe wants to achieve climate neutrality and energy sovereignty, it must stop ignoring geothermal. We need bold investment, regulatory reform, and a clear signal to industry: don’t let geothermal become a forgotten renewable. Grid failures, missed climate targets, deeper energy dependence — these are the risks Europe faces. It’s time to start drilling, before it’s too late.  Want to discover the next big thing in tech? Then take a trip to TNW Conference, where thousands of founders, investors, and corporate innovators will share their ideas. The event takes place on June 19–20 in Amsterdam and tickets are on sale now. Use the code TNWXMEDIA2025 at the checkout to get 30% off. Story by Siôn Geschwindt Siôn is a freelance science and technology reporter, specialising in climate and energy. From nuclear fusion breakthroughs to electric vehic (show all) Siôn is a freelance science and technology reporter, specialising in climate and energy. From nuclear fusion breakthroughs to electric vehicles, he's happiest sourcing a scoop, investigating the impact of emerging technologies, and even putting them to the test. He has five years of journalism experience and holds a dual degree in media and environmental science from the University of Cape Town, South Africa. When he's not writing, you can probably find Siôn out hiking, surfing, playing the drums or catering to his moderate caffeine addiction. You can contact him at: sion.geschwindt [at] protonmail [dot] com Get the TNW newsletter Get the most important tech news in your inbox each week. Also tagged with

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Alibaba Group has introduced QwenLong-L1, a new framework that enables large language models (LLMs) to reason over extremely long inputs. This development could unlock a new wave of enterprise applications that require models to understand and draw insights from extensive documents such as detailed corporate filings, lengthy financial statements, or complex legal contracts. The challenge of long-form reasoning for AI Recent advances in large reasoning models (LRMs), particularly through reinforcement learning (RL), have significantly improved their problem-solving capabilities. Research shows that when trained with RL fine-tuning, LRMs acquire skills similar to human “slow thinking,” where they develop sophisticated strategies to tackle complex tasks. However, these improvements are primarily seen when models work with relatively short pieces of text, typically around 4,000 tokens. The ability of these models to scale their reasoning to much longer contexts (e.g., 120,000 tokens) remains a major challenge. Such long-form reasoning requires a robust understanding of the entire context and the ability to perform multi-step analysis. “This limitation poses a significant barrier to practical applications requiring interaction with external knowledge, such as deep research, where LRMs must collect and process information from knowledge-intensive environments,” the developers of QwenLong-L1 write in their paper. The researchers formalize these challenges into the concept of “long-context reasoning RL.” Unlike short-context reasoning, which often relies on knowledge already stored within the model, long-context reasoning RL requires models to retrieve and ground relevant information from lengthy inputs accurately. Only then can they generate chains of reasoning based on this incorporated information.  Training models for this through RL is tricky and often results in inefficient learning and unstable optimization processes. Models struggle to converge on good solutions or lose their ability to explore diverse reasoning paths. QwenLong-L1: A multi-stage approach QwenLong-L1 is a reinforcement learning framework designed to help LRMs transition from proficiency with short texts to robust generalization across long contexts. The framework enhances existing short-context LRMs through a carefully structured, multi-stage process: Warm-up Supervised Fine-Tuning (SFT): The model first undergoes an SFT phase, where it is trained on examples of long-context reasoning. This stage establishes a solid foundation, enabling the model to ground information accurately from long inputs. It helps develop fundamental capabilities in understanding context, generating logical reasoning chains, and extracting answers. Curriculum-Guided Phased RL: At this stage, the model is trained through multiple phases, with the target length of the input documents gradually increasing. This systematic, step-by-step approach helps the model stably adapt its reasoning strategies from shorter to progressively longer contexts. It avoids the instability often seen when models are abruptly trained on very long texts. Difficulty-Aware Retrospective Sampling: The final training stage incorporates challenging examples from the preceding training phases, ensuring the model continues to learn from the hardest problems. This prioritizes difficult instances and encourages the model to explore more diverse and complex reasoning paths. QwenLong-L1 process Source: arXiv Beyond this structured training, QwenLong-L1 also uses a distinct reward system. While training for short-context reasoning tasks often relies on strict rule-based rewards (e.g., a correct answer in a math problem), QwenLong-L1 employs a hybrid reward mechanism. This combines rule-based verification, which ensures precision by checking for strict adherence to correctness criteria, with an “LLM-as-a-judge.” This judge model compares the semanticity of the generated answer with the ground truth, allowing for more flexibility and better handling of the diverse ways correct answers can be expressed when dealing with long, nuanced documents. Putting QwenLong-L1 to the test The Alibaba team evaluated QwenLong-L1 using document question-answering (DocQA) as the primary task. This scenario is highly relevant to enterprise needs, where AI must understand dense documents to answer complex questions.  Experimental results across seven long-context DocQA benchmarks showed QwenLong-L1’s capabilities. Notably, the QWENLONG-L1-32B model (based on DeepSeek-R1-Distill-Qwen-32B) achieved performance comparable to Anthropic’s Claude-3.7 Sonnet Thinking, and outperformed models like OpenAI’s o3-mini and Qwen3-235B-A22B. The smaller QWENLONG-L1-14B model also outperformed Google’s Gemini 2.0 Flash Thinking and Qwen3-32B.  Source: arXiv An important finding relevant to real-world applications is how RL training results in the model developing specialized long-context reasoning behaviors. The paper notes that models trained with QwenLong-L1 become better at “grounding” (linking answers to specific parts of a document), “subgoal setting” (breaking down complex questions), “backtracking” (recognizing and correcting their own mistakes mid-reasoning), and “verification” (double-checking their answers). For instance, while a base model might get sidetracked by irrelevant details in a financial document or get stuck in a loop of over-analyzing unrelated information, the QwenLong-L1 trained model demonstrated an ability to engage in effective self-reflection. It could successfully filter out these distractor details, backtrack from incorrect paths, and arrive at the correct answer. Techniques like QwenLong-L1 could significantly expand the utility of AI in the enterprise. Potential applications include legal tech (analyzing thousands of pages of legal documents), finance (deep research on annual reports and financial filings for risk assessment or investment opportunities) and customer service (analyzing long customer interaction histories to provide more informed support). The researchers have released the code for the QwenLong-L1 recipe and the weights for the trained models. Daily insights on business use cases with VB Daily If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for maximum ROI. Read our Privacy Policy Thanks for subscribing. Check out more VB newsletters here. An error occured.

While Europe races to phase out fossil fuels and electrify everything from cars to heating systems, it’s turning a blind eye to a reliable and proven source of clean energy lying right beneath our feet.  Geothermal energy offers exactly what the continent needs most: clean, local, always-on power. Yet, it only accounted for 0.2% of power generation on the continent in 2024. Something needs to change. The recent blackout in Spain, triggered by a failure in the high-voltage grid, serves as a warning shot. While solar and wind are vital pillars of decarbonisation, they’re variable by nature. Without steady, around-the-clock sources of electricity, Europe risks swapping one form of energy insecurity for another. A much bigger wake-up call came in 2022 when Russia launched a full-scale invasion of Ukraine. For years, European governments had built an energy system dependent on imports of natural gas. When that stack of cards shattered, it triggered an energy crisis that exposed the vulnerable underbelly of Europe’s power system.  The answer to these problems lies, in part, a few kilometres underground. According to the International Energy Agency, geothermal energy has the potential to power the planet 150 times over. But it’s not just about electricity — geothermal can also deliver clean, reliable heat. That makes it especially valuable in Europe, where millions of homes already rely on radiators and district heating systems, many of them still powered by natural gas. Geothermal plants also come with a smaller footprint. They require far less land than an equivalent solar farm or wind park. What’s more, the materials and infrastructure needed to build them — like drilling rigs and turbines — can be largely sourced locally. That’s a sharp contrast to solar panels and batteries, most of which are imported from China.   Geothermal energy is not theoretical. It doesn’t require scientific breakthroughs. We’ve been drilling wells and extracting energy from the Earth for centuries. The know-how exists, and so does the workforce. Decades of oil and gas exploration have built a deep bench of geologists, drillers, reservoir engineers, and project managers. Instead of letting this expertise fade, we can redeploy it to build geothermal plants. The infrastructure, such as drilling rigs, can also be repurposed for a cleaner cause. Geothermal could be the ultimate redemption arc for oil and gas. Sure, drilling deep isn’t cheap — yet. But a new crop of startups is rewriting the playbook. Armed with everything from plasma pulse drills to giant radiators, these companies could finally crack the cost barrier — and make geothermal available pretty much anywhere. Just as SpaceX disrupted a sclerotic rocket industry with its cheap launches, these startups are poised to succeed where the geothermal industry has failed.  All that’s missing is investment. While billions are being funnelled into high-risk technologies like fusion or nuclear fission reactors, funding for geothermal tech is minuscule in comparison, especially in Europe. Yet, unlike those technologies, geothermal is ready right now.   If Europe wants to achieve climate neutrality and energy sovereignty, it must stop ignoring geothermal. We need bold investment, regulatory reform, and a clear signal to industry: don’t let geothermal become a forgotten renewable. Grid failures, missed climate targets, deeper energy dependence — these are the risks Europe faces. It’s time to start drilling, before it’s too late.  Want to discover the next big thing in tech? Then take a trip to TNW Conference, where thousands of founders, investors, and corporate innovators will share their ideas. The event takes place on June 19–20 in Amsterdam and tickets are on sale now. Use the code TNWXMEDIA2025 at the checkout to get 30% off. Story by Siôn Geschwindt Siôn is a freelance science and technology reporter, specialising in climate and energy. From nuclear fusion breakthroughs to electric vehic (show all) Siôn is a freelance science and technology reporter, specialising in climate and energy. From nuclear fusion breakthroughs to electric vehicles, he's happiest sourcing a scoop, investigating the impact of emerging technologies, and even putting them to the test. He has five years of journalism experience and holds a dual degree in media and environmental science from the University of Cape Town, South Africa. When he's not writing, you can probably find Siôn out hiking, surfing, playing the drums or catering to his moderate caffeine addiction. You can contact him at: sion.geschwindt [at] protonmail [dot] com Get the TNW newsletter Get the most important tech news in your inbox each week. Also tagged with

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Alibaba Group has introduced QwenLong-L1, a new framework that enables large language models (LLMs) to reason over extremely long inputs. This development could unlock a new wave of enterprise applications that require models to understand and draw insights from extensive documents such as detailed corporate filings, lengthy financial statements, or complex legal contracts. The challenge of long-form reasoning for AI Recent advances in large reasoning models (LRMs), particularly through reinforcement learning (RL), have significantly improved their problem-solving capabilities. Research shows that when trained with RL fine-tuning, LRMs acquire skills similar to human “slow thinking,” where they develop sophisticated strategies to tackle complex tasks. However, these improvements are primarily seen when models work with relatively short pieces of text, typically around 4,000 tokens. The ability of these models to scale their reasoning to much longer contexts (e.g., 120,000 tokens) remains a major challenge. Such long-form reasoning requires a robust understanding of the entire context and the ability to perform multi-step analysis. “This limitation poses a significant barrier to practical applications requiring interaction with external knowledge, such as deep research, where LRMs must collect and process information from knowledge-intensive environments,” the developers of QwenLong-L1 write in their paper. The researchers formalize these challenges into the concept of “long-context reasoning RL.” Unlike short-context reasoning, which often relies on knowledge already stored within the model, long-context reasoning RL requires models to retrieve and ground relevant information from lengthy inputs accurately. Only then can they generate chains of reasoning based on this incorporated information.  Training models for this through RL is tricky and often results in inefficient learning and unstable optimization processes. Models struggle to converge on good solutions or lose their ability to explore diverse reasoning paths. QwenLong-L1: A multi-stage approach QwenLong-L1 is a reinforcement learning framework designed to help LRMs transition from proficiency with short texts to robust generalization across long contexts. The framework enhances existing short-context LRMs through a carefully structured, multi-stage process: Warm-up Supervised Fine-Tuning (SFT): The model first undergoes an SFT phase, where it is trained on examples of long-context reasoning. This stage establishes a solid foundation, enabling the model to ground information accurately from long inputs. It helps develop fundamental capabilities in understanding context, generating logical reasoning chains, and extracting answers. Curriculum-Guided Phased RL: At this stage, the model is trained through multiple phases, with the target length of the input documents gradually increasing. This systematic, step-by-step approach helps the model stably adapt its reasoning strategies from shorter to progressively longer contexts. It avoids the instability often seen when models are abruptly trained on very long texts. Difficulty-Aware Retrospective Sampling: The final training stage incorporates challenging examples from the preceding training phases, ensuring the model continues to learn from the hardest problems. This prioritizes difficult instances and encourages the model to explore more diverse and complex reasoning paths. QwenLong-L1 process Source: arXiv Beyond this structured training, QwenLong-L1 also uses a distinct reward system. While training for short-context reasoning tasks often relies on strict rule-based rewards (e.g., a correct answer in a math problem), QwenLong-L1 employs a hybrid reward mechanism. This combines rule-based verification, which ensures precision by checking for strict adherence to correctness criteria, with an “LLM-as-a-judge.” This judge model compares the semanticity of the generated answer with the ground truth, allowing for more flexibility and better handling of the diverse ways correct answers can be expressed when dealing with long, nuanced documents. Putting QwenLong-L1 to the test The Alibaba team evaluated QwenLong-L1 using document question-answering (DocQA) as the primary task. This scenario is highly relevant to enterprise needs, where AI must understand dense documents to answer complex questions.  Experimental results across seven long-context DocQA benchmarks showed QwenLong-L1’s capabilities. Notably, the QWENLONG-L1-32B model (based on DeepSeek-R1-Distill-Qwen-32B) achieved performance comparable to Anthropic’s Claude-3.7 Sonnet Thinking, and outperformed models like OpenAI’s o3-mini and Qwen3-235B-A22B. The smaller QWENLONG-L1-14B model also outperformed Google’s Gemini 2.0 Flash Thinking and Qwen3-32B.  Source: arXiv An important finding relevant to real-world applications is how RL training results in the model developing specialized long-context reasoning behaviors. The paper notes that models trained with QwenLong-L1 become better at “grounding” (linking answers to specific parts of a document), “subgoal setting” (breaking down complex questions), “backtracking” (recognizing and correcting their own mistakes mid-reasoning), and “verification” (double-checking their answers). For instance, while a base model might get sidetracked by irrelevant details in a financial document or get stuck in a loop of over-analyzing unrelated information, the QwenLong-L1 trained model demonstrated an ability to engage in effective self-reflection. It could successfully filter out these distractor details, backtrack from incorrect paths, and arrive at the correct answer. Techniques like QwenLong-L1 could significantly expand the utility of AI in the enterprise. Potential applications include legal tech (analyzing thousands of pages of legal documents), finance (deep research on annual reports and financial filings for risk assessment or investment opportunities) and customer service (analyzing long customer interaction histories to provide more informed support). The researchers have released the code for the QwenLong-L1 recipe and the weights for the trained models. Daily insights on business use cases with VB Daily If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for maximum ROI. Read our Privacy Policy Thanks for subscribing. Check out more VB newsletters here. An error occured.

Large language models (LLMs), trained on datasets with billions of tokens, can generate high-quality content. They’re the backbone for many of the most popular AI applications, including chatbots, assistants, code generators and much more. One of today’s most accessible ways to work with LLMs is with AnythingLLM, a desktop app built for enthusiasts who want an all-in-one, privacy-focused AI assistant directly on their PC. With new support for NVIDIA NIM microservices on NVIDIA GeForce RTX and NVIDIA RTX PRO GPUs, AnythingLLM users can now get even faster performance for more responsive local AI workflows. What Is AnythingLLM? AnythingLLM is an all-in-one AI application that lets users run local LLMs, retrieval-augmented generation (RAG) systems and agentic tools. It acts as a bridge between a user’s preferred LLMs and their data, and enables access to tools (called skills), making it easier and more efficient to use LLMs for specific tasks like: Question answering: Getting answers to questions from top LLMs — like Llama and DeepSeek R1 — without incurring costs. Personal data queries: Use RAG to query content privately, including PDFs, Word files, codebases and more. Document summarization: Generating summaries of lengthy documents, like research papers. Data analysis: Extracting data insights by loading files and querying it with LLMs. Agentic actions: Dynamically researching content using local or remote resources, running generative tools and actions based on user prompts. AnythingLLM can connect to a wide variety of open-source local LLMs, as well as larger LLMs in the cloud, including those provided by OpenAI, Microsoft and Anthropic. In addition, the application provides access to skills for extending its agentic AI capabilities via its community hub. With a one-click install and the ability to launch as a standalone app or browser extension — wrapped in an intuitive experience with no complicated setup required — AnythingLLM is a great option for AI enthusiasts, especially those with GeForce RTX and NVIDIA RTX PRO GPU-equipped systems. RTX Powers AnythingLLM Acceleration GeForce RTX and NVIDIA RTX PRO GPUs offer significant performance gains for running LLMs and agents in AnythingLLM — speeding up inference with Tensor Cores designed to accelerate AI. AnythingLLM runs LLMs with Ollama for on-device execution accelerated through Llama.cpp and ggml tensor libraries for machine learning. Ollama, Llama.cpp and GGML are optimized for NVIDIA RTX GPUs and the fifth-generation Tensor Cores. Performance on GeForce RTX 5090 is 2.4X compared to an Apple M3 Ultra. GeForce RTX 5090 delivers 2.4x faster LLM inference in AnythingLLM than Apple M3 Ultra on both Llama 3.1 8B and DeepSeek R1 8B. As NVIDIA adds new NIM microservices and reference workflows — like its growing library of AI Blueprints — tools like AnythingLLM will unlock even more multimodal AI use cases. AnythingLLM — Now With NVIDIA NIM AnythingLLM recently added support for NVIDIA NIM microservices — performance-optimized, prepackaged generative AI models that make it easy to get started with AI workflows on RTX AI PCs with a streamlined API. NVIDIA NIMs are great for developers looking for a quick way to test a Generative AI model in a workflow. Instead of having to find the right model, download all the files and figure out how to connect everything, they provide a single container that has everything you need. And they can run both on Cloud and PC, making it easy to prototype locally and then deploy on the cloud. By offering them within AnythingLLM’s user-friendly UI, users have a quick way to test them and experiment with them. And then they can either connect them to their workflows with AnythingLLM, or leverage NVIDIA AI Blueprints and NIM documentation and sample code to plug them directly to their apps or projects. Explore the wide variety of NIM microservices available to elevate AI-powered workflows, including language and image generation, computer vision and speech processing. Each week, the RTX AI Garage blog series features community-driven AI innovations and content for those looking to learn more about NIM microservices and AI Blueprints, as well as building AI agents, creative workflows, digital humans, productivity apps and more on AI PCs and workstations.  Plug in to NVIDIA AI PC on Facebook, Instagram, TikTok and X — and stay informed by subscribing to the RTX AI PC newsletter. Follow NVIDIA Workstation on LinkedIn and X. See notice regarding software product information.