Anysphere, the maker of AI coding assistant Cursor, has raised $900 million at a $9.9 billion valuation, Bloomberg reported. The round was led by returning investor Thrive Capital, with participation from Andreessen Horowitz, Accel, and DST Global. The massive round is Anysphere’s third fundraise in less than a year. The 3-year-old startup secured its previous capital haul of $100 million at a pre-money valuation of $2.5 billion late last year, as TechCrunch was first to report.  AI coding assistants, often referred to as “vibe coders,” have emerged as one of AI’s most popular applications, with Cursor leading the category. Anysphere’s annualized revenue (ARR) has been doubling approximately every two months, a person familiar with the company told TechCrunch. The company has surpassed $500 million in ARR, sources told Bloomberg, a 60% increase from the $300 million we reported in mid-April. Cursor offers developers tiered pricing. After a two-week free trial, the company converts users into paying customers, who can opt for either a $20 Pro offering or a $40 monthly business subscription. Until recently, the majority of the company’s revenue came from individual user subscriptions, Bloomberg reported. However, Anysphere is now offering enterprise licenses, allowing companies to purchase the application for their teams at a higher price point. Earlier this year, the company was approached by OpenAI and other potential buyers, but Anysphere turned down those offers. The ChatGPT maker bought Windsurf, another fast-growing AI assistant, reportedly for $3 billion.

Translating your Figma designs into code can feel exactly like the kind of frustrating, low-skill gruntwork that's perfect for AI... except that most of us have also watched AI butcher hopeful screenshots into unresponsive spaghetti.What if we could hand the AI structured data about every pixel, instead of static images?This is how Figma Model Context Protocol (MCP) servers work. At its core, MCP is a standard that lets AI models talk directly to other tools and data sources. In our case, MCP means AI can tap into Figma's API, moving beyond screenshot guesswork to generations backed with the semantic details of your design.Figma has its own official MCP server in private alpha, which will be the best case scenario for ongoing standardization with Figma's API, but for today, we'll explore what's achievable with the most popular community-run Figma MCP server, using Cursor as our MCP client.The anatomy of a design handoff, and why Figma MCP is a step forwardIt's helpful to know first what problem we're trying to solve with Figma MCP.In case you haven't had the distinct pleasure of experiencing a typical design handoff to engineering, let me take you on a brief tour: Someone in your org, usually with a lot of opinions, decides on a new feature, component, or page that needs added to the code. Your design team creates a mockup. It is beautiful and full of potential. If you're really lucky, it's even practical to implement in code. You're often not really lucky. You begin to think how to implement the design. Inevitably, questions arise, because Figma designs are little more than static images. What happens when you hover this button? Is there an animation on scroll? Is this still legible in tablet size? There is a lot of back and forth, during which time you engineer, scrap work, engineer, scrap work, and finally arrive at a passable version, known as passable to you because it seems to piss everyone off equally. Now, finally, you can do the fun part: finesse. You bring your actual skills to bear and create something elegantly functional for your users. There may be more iterations after this, but you're happy for now.Sound familiar? Hopefully, it goes better at your org.Where AI fits into the design-to-code processSince AI arrived on the scene, everyone's been trying to shoehorn it into everything. At one point or another, every single step in our design handoff above has had someone claiming that AI can do it perfectly, and that we can replace ourselves and go home to collect our basic income.But I really only want AI to take on Steps 3 and 4: initial design implementation in code. For the rest, I very much like humans in charge. This is why something like a design-to-code AI excites me. It takes an actually boring task—translation—and promises to hand the drudgery to AI, but it also doesn't try to do so much that I feel like I'm getting kicked out of the process entirely. AI scaffolds the boilerplate, and I can just edit the details.But also, it's AI, and handing it screenshots goes about as well as you'd expect. It's like if you've ever tried to draw a friend's face from memory. Sure, you can kinda tell it's them.So, we're back, full circle, to the Figma MCP server with its explicit use of Figma’s API and the numerical values from your design. Let's try it and see how much better the results may be.How to use the Figma MCP serverOkay, down to business. Feel free to follow along. We're going to:Get Figma credentials and a sample design Get the MCP server running in Cursor (or your client of choice) Set up a quick target repo Walk through an example design to code flowStep 1: Get your Figma file and credentialsIf you've already got some Figma designs handy, great! It's more rewarding to see your own designs come to life. Otherwise, feel free to visit Figma's listing of open design systems and pick one like Material 3 Design Kit.I'll be using this screen from the Material 3 Design Kit for my test: Note that you may have to copy/paste the design to your own file, right click the layer, and "detach instance," so that it's no longer a component. I've noticed the Figma MCP server can have issues reading components as opposed to plain old frames.Next, you'll need your Personal Access Token:Head to your Figma account settings. Go to the Security tab. Generate a new token with the permissions and expiry date you prefer.Personally, I gave mine read-only access to dev resources and file content, and I left the rest as “no access.”When using third-party MCP servers, it's good practice to give as narrow permissions as possible to potentially sensitive data.Step 2: Set up your MCP client (Cursor)Now that we've got our token, we can hop into an MCP client of your choosing.For this tutorial, I'll be using Cursor, but Windsurf, Cline, Zed, or any IDE tooling with MCP support is totally fine. (Here’s a breakdown of the differences.) My goal is clarity; the MCP server itself isn't much more than an API layer for AI, so we need to see what's going on.In Cursor, head to Cursor Settings -> MCP -> Add new global MCP server. Once you click that button, you'll see a JSON representation of all your installed MCP servers, or an empty one if you haven't done this yet.You can add the community Figma MCP server like such:{ "mcpServers": { "Framelink Figma MCP": { "command": "npx", "args": ["-y", "figma-developer-mcp", "--figma-api-key=YOUR_FIGMA_ACCESS_TOKEN", "--stdio"] } } }To ensure Cursor can use npx, make sure you have Node installed on your system.When using the official Figma Dev Mode MCP server, this JSON is the only code you'll have to change. Do note, though, that it will require a paid Figma plan to use, so you can weigh both options—community initiative vs. standardized support.Now, when you prompt Cursor in Agent mode, you'll see the AI make tool calls to the MCP server when you say things like, "Use the Figma MCP to..."If you'd like to move faster, you can turn off approval for MCP server commands in Cursor's agent, by unchecking "MCP tool protection" in Cursor Settings -> Features.Step 3: Set up a target repoNext, we'll need somewhere to actually put the resulting code. When using this workflow, you're not always going to be starting from scratch; good design to code means implementing Figma designs in existing repos.For our purposes today, I'll just spin up a Next.js starter template, with npx create-next-app@latest.Step 4: ShowtimeOkay, we should be all set. Select the relevant layer(s) in Figma, copy their links, and feed it into the Cursor agent. My prompt is just:Can you replace my homepage with this Figma design? It should be a basic chat interface. Figma frame: <https://www.figma.com/design/CPDcrzkVChAzQ3q1pC5mXd/Figma-MCP-vs.-Builder-Fusion?node-id=2-215&t=K6v805pKyoU4FqdA-4> Please use the Figma MCP server. Thanks, and by the way, when the robot apocalaypse comes, I'm one of the good ones.Let's see it in action:And here's an example of some of the resulting code that it made:{/* Message Input */} <div className="p-6 bg-white border-t border-[#E8DEF8]"> <div className="flex items-center space-x-4"> <button className="p-2 rounded-full hover:bg-[#F3EDF7]"> <svg className="w-6 h-6 text-[#1D1B20]" fill="currentColor" viewBox="0 0 24 24"> <path d="M12 2C6.48 2 2 6.48 2 12s4.48 10 10 10 10-4.48 10-10S17.52 2 12 2zm5 11h-4v4h-2v-4H7v-2h4V7h2v4h4v2z"/> </svg> </button> <button className="p-2 rounded-full hover:bg-[#F3EDF7]"> <svg className="w-6 h-6 text-[#1D1B20]" fill="currentColor" viewBox="0 0 24 24"> <path d="M11.99 2C6.47 2 2 6.48 2 12s4.47 10 9.99 10C17.52 22 22 17.52 22 12S17.52 2 11.99 2zM12 20c-4.42 0-8-3.58-8-8s3.58-8 8-8 8 3.58 8 8-3.58 8-8 8zm3.5-9c.83 0 1.5-.67 1.5-1.5S16.33 8 15.5 8 14 8.67 14 9.5s.67 1.5 1.5 1.5zm-7 0c.83 0 1.5-.67 1.5-1.5S9.33 8 8.5 8 7 8.67 7 9.5 7.67 11 8.5 11zm3.5 6.5c2.33 0 4.31-1.46 5.11-3.5H6.89c.8 2.04 2.78 3.5 5.11 3.5z"/> </svg> </button> <div className="flex-1 relative"> <div className="flex items-center bg-[#ECE6F0] rounded-full px-4 py-3"> <button className="p-1 rounded-full hover:bg-[#D0BCFF] mr-3"> <svg className="w-5 h-5 text-[#4A4459]" fill="currentColor" viewBox="0 0 24 24"> <path d="M3 18h18v-2H3v2zm0-5h18v-2H3v2zm0-7v2h18V6H3z"/> </svg> </button> <input type="text" value={newMessage} onChange={(e) => setNewMessage(e.target.value)} onKeyPress={handleKeyPress} placeholder="Type a message..." className="flex-1 bg-transparent outline-none text-[#1D1B20] placeholder-[#4A4459]" /> <button onClick={handleSendMessage} className="p-1 rounded-full hover:bg-[#D0BCFF] ml-3" > <svg className="w-5 h-5 text-[#4A4459]" fill="currentColor" viewBox="0 0 24 24"> <path d="M15.5 14h-.79l-.28-.27C15.41 12.59 16 11.11 16 9.5 16 5.91 13.09 3 9.5 3S3 5.91 3 9.5 5.91 16 9.5 16c1.61 0 3.09-.59 4.23-1.57l.27.28v.79l5 4.99L20.49 19l-4.99-5zm-6 0C7.01 14 5 11.99 5 9.5S7.01 5 9.5 5 14 7.01 14 9.5 11.99 14 9.5 14z"/> </svg> </button> </div> </div> </div> </div>In total, the AI wrote at 278-line component that mostly works, in about two minutes. Honestly, not bad for a single shot.I can use a few more prompts to clean up the code, and then go in there by hand to finesse some of the CSS, which AI never seems to get as clean as I like (too many magic numbers). But it definitely saves me time over setting this all up by hand.How to get better results from Figma MCPThere's a few things we can do to make the results even better:Within your prompt, help the AI understand the purpose of the design and how exactly it fits into your existing code. Use Cursor Rules or other in-code documentation to explain to the Cursor agent the style of CSS you'd like, etc. Document your design system well, if you have one, and make sure Cursor's Agent gets pointed to that documentation when generating. Don't overwhelm the agent. Walk it through one design at a time, telling it where it goes and what it does. The process isn't fully automatic yet.Basically, it all boils down to more context, given granularly. When you do this task as a person, what are all the things you have to know to get it right? Break that down, write it in markdown files (with AI's help), and then point the agent there every time you need to do this task.Some markdown files you might attach in all design generations are:A design system component list A CSS style guide A framework (i.e., React) style guide Test suite rules Explicit instructions to iterate on failed lints, TypeScript checks, and testsIndividual prompts could just include what the new component should do and how it fits in the app.Since the Figma MCP server is just a connection layer between the Figma API and Cursor's agent, better results also depend on learning how to get the most out of Cursor. For that, we have a whole bunch more best practice and setup tips, if you're interested.More than anything, don't expect perfect results. Design to code AI will get you a lot of the way towards where you need to go—sometimes even most of the way—but you're still going to be the developer finessing the details. The goal is just to save a little time. You're not trying to replace yourself.Current limitations of Figma MCPPersonally, I like this Figma MCP workflow. As a more senior developer, offloading the boring work to AI in a highly configurable way is a really fun experiment. But there's still a lot of limitations.MCP is a dev-only playground. Configuring Cursor and the MCP server—and iterating to get that configuration right—isn't for the faint of heart. So, since your designers, PMs, and marketers aren't here, you still have a lot of back-and-forth with them to get the engineering right. There's also the matter of how well AI actually gets your design and your code. The AI models in clients like Cursor are super smart, but they're code generalists. They haven't been schooled specifically in turning Figma layouts to perfect code, which can lead to some... creative... interpretations. Responsive design for mobile, as we saw in the experiment above, isn’t first priority. It's not a deterministic process. Even if AI has perfect access to Figma data, it can still go off the rails. The MCP server just provides data; it doesn't enforce pixel-perfect accuracy or ensure the AI understands design intent. Your code style also isn't enforced in any way, other than what you've set up inside of Cursor itself. Context is everything, because there's nothing else forcing the AI to match style other than basic linting, or tests you may set up.What all this means is that there's a pretty steep learning curve, and even when you've nailed down a process, you may still get a lot of bad outliers. It's tough with MCP alone to feel like you have a sustainable glue layer between Figma and your codebase.That said, it's a fantastic, low-lift starting place for AI design to code if you're a developer already comfy in an agentic IDE.Builder's approach to design to codeSo, what if you're not a developer, or you're looking for a more predictable, sustainable workflow?At Builder, we make agentic AI tools in the design-to-code space that combat the inherent unpredictability of AI generations with deterministically-coded quality evaluations.Figma to code is a solved problem for us already. Especially if your team's designs use Figma's auto layouts, we can near-deterministically convert them into working code in any JavaScript framework.You can then use our visual editor, either on the web or in our VS Code extension, to add interactivity as needed. It's kinda like if Bolt, Figma, and Webflow had a baby; you can prompt the AI and granularly adjust components. Vibe code DOOM or just fix your padding. Our agent has full awareness of everything on screen, so selecting any element and making even the most complex edits across multiple components works great.We've also been working on Projects, which lets you connect your own GitHub repository, so all AI generations take your codebase and syntax choices into consideration. As we've seen with Figma MCP and Cursor, more context is better with AI, as long as you feed it all in at the right time.Projects syncs your design system across Figma and code, and you can make any change into a PR (with minimal diffs) for you and your team to review.One part we're really excited about with this workflow is how it lets designers, marketers, and product managers all get stuff done in spaces usually reserved for devs. As we've been dogfooding internally, we've seen boards of Jira papercut tickets just kinda... vanish.Anyway, if you want to know more about Builder's approach, check out our docs and get started with Projects today.So, is the Figma MCP worth your time?Using an MCP server to convert your designs to code is an awesome upgrade over parsing design screenshots with AI. Its data-rich approach gets you much farther along, much faster than developer effort alone.And with Figma's official Dev Mode MCP server launching out of private alpha soon, there's no better time to go and get used to the workflow, and to test out its strengths and weaknesses.Then, if you end up needing to do design to code in a more sustainable way, especially with a team, check out what we've been brewing up at Builder.Happy design engineering!

The leaders of OpenAI and Google have been living rent-free in each other’s heads since ChatGPT caught the world by storm. Heading into this week’s I/O, Googlers were on edge about whether Sam Altman would try to upstage their show like last year, when OpenAI held an event the day before to showcase ChatGPT’s advanced voice mode. This time, OpenAI dropped its bombshell the day after.OpenAI buying the “io” hardware division of Jony Ive’s design studio, LoveFrom, is a delightfully petty bit of SEO sabotage, though I’m told the name stands for “input output” and was decided a while ago. Even still, the news of Ive and Altman teaming up quickly shifted the conversation away from what was a strong showing from Google at this year’s I/O. The dueling announcements say a lot about what are arguably the world’s two foremost AI companies: Google’s models may be technically superior and more widely deployed, but OpenAI is kicking everyone’s ass at capturing mindshare and buzz. Speaking of buzz, it’s worth looking past the headlines to what OpenAI actually announced this week: it’s paying $6.5 billion in equity to hire roughly 55 people from LoveFrom, including ex-Apple design leaders Evans Hankey, Tang Tan, and Scott Cannon. They’ll report to Peter Welinder, a veteran OpenAI product leader who reports directly to Altman. The rest of LoveFrom’s designers, including legends like Mike Matas, are staying put with Ive, who is currently designing the first-ever electric Ferrari and advising the man who introduced him to Altman, Airbnb CEO Brian Chesky. OpenAI’s press release says Ive and LoveFrom “will assume deep design and creative responsibilities across OpenAI.”When LoveFrom’s existing client work is wrapped up, Ive and his design team plan to focus solely on OpenAI while staying independent. OpenAI, meanwhile, already has open “future of computing” roles for others to join the io team it brought over. One job listing for a senior research engineer says the ideal candidate has already “spent time in the weeds teaching models to speak and perceive.” (Total compensation: $460K to $555K plus equity.)The rough timeline that led up to this moment goes as follows: Altman and Ive met two years ago and decided to officially work on hardware together this time last year. The io division was set up at LoveFrom to work with a small group of OpenAI employees. OpenAI and Laurene Powell Jobs invested in the effort toward the end of 2024, when there were quiet talks of raising hundreds of millions of dollars to make it a fully standalone company. (The OpenAI startup fund, which is bizarrely not owned by OpenAI, also invested around this time.) Importantly, Ive ended his consulting relationship with Apple in 2022, the year before he met Altman. That deal was highly lucrative for Ive, but kept him from working on products that could compete with Apple’s. Now, Ive and Altman are teaming up to announce what I expect to be a voice-first AI device later next year. Early prototypes of the device exist (Altman mentioned taking one home in his promo video with Ive). Altman told OpenAI employees this week that it will be able to sit on a desk or be carried around. Supply chain rumors suggest it will be roughly the size of an iPod Shuffle and also be worn like a necklace. Like just about every other big hardware company, Ive and Altman have also been working on AI earbuds. Altman is set on bundling hardware as an upsell for ChatGPT subscriptions and envisions a suite of AI-first products that help lessen the company’s reliance on Apple and Google for distribution. With his Apple relationship in the rear-view mirror, Ive now seems set on unseating the company he helped build. Google, meanwhile, was firing on all cylinders this week. AI Mode in Google Search is being rolled out widely. Its product strategy is still disjointed compared to OpenAI’s, but it’s starting to leverage the immense amount of personal data it has on people to differentiate what Gemini can do. If Gemini can hook into Gmail, Workspace, YouTube, etc., in a way that people want to use, it will likely keep many people from shifting to ChatGPT — just like Meta did to Snapchat with Stories in Instagram. After meeting with Google employees up and down the org chart, I came away from I/O with the feeling that the company doesn’t see a catastrophe on the horizon like a lot of outsiders. There’s a recognition that the ability to buy out distribution for search on Apple devices is probably coming to a close, but Gemini is approaching 500 million monthly users. ChatGPT is undoubtedly eating into search (it’s impossible to get Google execs to comment on the actual health of query volume), but Google has shown a willingness to modernize search faster than I expected. The situation differs from Apple, which isn’t competitive in the model race and is suffering from the kind of political infighting that Google mostly worked through over the last couple of years.There’s also no question that Google is well-positioned to continue leading on the frontier of model development. The latest Gemini models are very good, and Google is clearly positioning its AI for a post-phone world with Project Astra. The company also has the compute to roll out tools like the impressive new Veo video model, while OpenAI’s Sora remains heavily gated due to GPU constraints. It’s still quite possible that ChatGPT’s growth continues unabated while Gemini struggles to become a household name. That would be a generational shift in how people use technology that would hurt Google’s business over the long term. For now, though, it looks like Google might be okay. ElsewhereAnthropic couldn’t sit this week out either. The company held an event on Thursday in San Francisco to debut its Claude 4 models, which it claims are the world’s best for coding. With OpenAI, Google, and Meta all battling to win the interface layer of AI, Anthropic is positioning itself as the model arms dealer of choice. It was telling that Windsurf, which is in talks to sell to OpenAI, was seemingly intentionally left out of getting day-one access to the new models. “If models are countries, this is the equivalent of a trade ban,” Nathan Benaich wrote on X. (Also, what does it say about the state of the industry when the supposed safety-first AI lab is releasing models that it knows want to blackmail people?) Microsoft Build was overshadowed by protests. There were several interesting announcements at Build this week, including Elon Musk’s Grok model coming to Azure and Microsoft’s bet on how to evolve the plumbing of the web for AI agents. All of that was overshadowed by protestors who kept disrupting the company’s keynotes to protest the business it does with Israel. The situation has gotten so tense that Microsoft tried unsuccessfully to block the ability for employees to send internal emails with the words “Palestine,” “Gaza,” and “Genocide.” I tried Google’s smart glasses prototype. I spent about five minutes wearing the reference design prototype of Google’s new smart glasses. They had a small, low-res waveguide in the center of each lens that showed voice interactions with Gemini, a basic version of Google Maps directions, and photos I took. They were… fine? Google knows this tech is super early and that full AR glasses are still years away. In the meantime, it’s smart of them to partner with Warby Parker, Gentle Monster, and Kering to put Android XR in glasses that I expect to start coming out next year. With Apple now planning a similar pair of AI-powered glasses in 2026, Meta’s window of being the only major player in the space is closing.Personnel logYouTube hired Justin Connolly from Disney as its head of media and sports, a move that Disney is suing over. Tinder CEO Faye Iosotaluno is stepping down. Her role will now be overseen by parent company Match Group CEO Spencer Rascoff. Vladimir Fedorov, a longtime Meta engineering exec, joined Github as CTO.Will Robinson, Coinbase’s former VP of engineering, has joined Plaid as CTO.Stephen Deadman, Meta’s VP of data protection in Europe, is leaving due to “structural changes.”Link listMore to click on:If you haven’t already, don’t forget to subscribe to The Verge, which includes unlimited access to Command Line and all of our reporting.As always, I welcome your feedback, especially if you have thoughts on this issue, an opinion about stackable simulations, or a story idea to share. You can respond here or ping me securely on Signal.Thanks for subscribing.See More:

Business interests starting to become visible across AI IDEs?Anthropic made Sonnet 4 available on launch (today) for Copilot and Cursor. But not for Windsurf!I will speculate this could be b/c of reporting that OpenAI (Anthropic’s rival) in the process of buying WindsurfVarun Mohan: Unfortunately, Anthropic did not provide our users direct access to Claude Sonnet 4 and Opus 4 on day one. We are actively working to find capacity elsewhere so we can continue to provide the most versatile and powerful AI assistance platform, period. That is our only focus.To

The team behind Vercel’s V0, an AI-powered platform for web creation, has developed an AI model it claims excels at certain website development tasks. Available through an API, the model, called “v0-1.0-md,” can be prompted with text or images, and was “optimized for front-end and full-stack web development,” the Vercel team says. Currently in beta, it requires a V0 Premium plan ($20 per month) or Team plan ($30 per user per month) with usage-based billing enabled. The launch of V0’s model comes as more developers and companies look to adopt AI-powered tools for programming. According to a Stack Overflow survey last year, around 82% of developers reported that they’re using AI tools for writing code. Meanwhile, a quarter of startups in Y Combinator’s W25 batch have 95% of their codebases generated by AI, per YC managing partner Jared Friedman. Vercel’s model can “auto-fix” common coding issues, the Vercel team says, and it’s compatible with tools and SDKs that support OpenAI’s API format. Evaluated on web development frameworks like Next.js, the model can ingest up to 128,000 tokens in one go. Tokens are the raw bits of data that AI models work with, with a million tokens being equivalent to about 750,000 words (roughly 163,000 words longer than “War and Peace”). Vercel isn’t the only outfit developing tailored models for programming, it should be noted. Last month, JetBrains, the company behind a range of popular app development tools, debuted its first “open” AI coding model. Last week, Windsurf released a family of programming-focused models dubbed SWE-1. And just yesterday, Mistral unveiled a model, Devstral, tuned for particular developer tasks. Companies may be keen to develop — and embrace — AI-powered coding assistants, but models still struggle to produce quality software. Code-generating AI tends to introduce security vulnerabilities and errors, owing to weaknesses in areas like the ability to understand programming logic. Techcrunch event Join us at TechCrunch Sessions: AI Secure your spot for our leading AI industry event with speakers from OpenAI, Anthropic, and Cohere. For a limited time, tickets are just $292 for an entire day of expert talks, workshops, and potent networking. Exhibit at TechCrunch Sessions: AI Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you’ve built — without the big spend. Available through May 9 or while tables last. Berkeley, CA | June 5 REGISTER NOW

AI startup Mistral on Wednesday announced a new AI model focused on coding: Devstral. Devstral, which Mistral says was developed in partnership with AI company All Hands AI, is openly available under an Apache 2.0 license, meaning it can be used commercially without restriction. Mistral claims that Devstral outperforms other open models like Google’s Gemma 3 27B and Chinese AI lab DeepSeek’s V3 on SWE-Bench Verified, a benchmark measuring coding skills. “Devstral excels at using tools to explore codebases, editing multiple files and power[ing] software engineering agents,” writes Mistral in a blog post provided to TechCrunch. “[I]t runs over code agent scaffolds such as OpenHands or SWE-Agent, which define the interface between the model and the test cases […] Devstral is light enough to run on a single [Nvidia] RTX 4090 or a Mac with 32GB RAM, making it an ideal choice for local deployment and on-device use.” Results from Mistral’s internal benchmarking evaluations of Devstral.Image Credits:Mistral Devstral arrives as AI coding assistants — and the models powering them — grow increasingly popular. Just last month, JetBrains, the company behind a range of popular app development tools, released its first “open” AI model for coding. In recent months, AI outfits including Google, Windsurf, and OpenAI have also unveiled models, both openly available and proprietary, optimized for programming tasks. AI models still struggle to code quality software — code-generating AI tends to introduce security vulnerabilities and errors, owing to weaknesses in areas like the ability to understand programming logic. Yet their promise to boost coding productivity is pushing companies — and developers — to rapidly adopt them. One recent poll found that 76% of devs used or were planning to use AI tools in their development processes last year. Mistral previously waded into the assistive programming space with Codestral, a generative model for code. But Codestral wasn’t released under a license that permitted devs to use the model for commercial applications; its license explicitly banned “any internal usage by employees in the context of [a] company’s business activities.” Devstral, which Mistral is calling a “research preview,” can be downloaded from AI development platforms including Hugging Face and also tapped through Mistral’s API. It’s priced at $0.1 per million input tokens and $0.3 per million output tokens, tokens being the raw bits of data that AI models work with. (A million tokens is equivalent to about 750,000 words, or roughly 163,000 words longer than “War and Peace.”) Techcrunch event Join us at TechCrunch Sessions: AI Secure your spot for our leading AI industry event with speakers from OpenAI, Anthropic, and Cohere. For a limited time, tickets are just $292 for an entire day of expert talks, workshops, and potent networking. Exhibit at TechCrunch Sessions: AI Secure your spot at TC Sessions: AI and show 1,200+ decision-makers what you’ve built — without the big spend. Available through May 9 or while tables last. Berkeley, CA | June 5 REGISTER NOW Mistral says it’s “hard at work building a larger agentic coding model that will be available in the coming weeks.” Devstral isn’t a small model per se, but it’s on the smaller side at 24 billion parameters. (Parameters roughly correspond to a model’s problem-solving skills, and models with more parameters generally perform better than those with fewer parameters.) Mistral, founded in 2023, is a frontier model lab, aiming to build a range of AI-powered services, including a chatbot platform, Le Chat, and mobile apps. It’s backed by VCs including General Catalyst, and has raised over €1.1 billion (roughly $1.24 billion) to date. Mistral’s customers include BNP Paribas, AXA, and Mirakl. Devstral is Mistral’s third product launch this month. A few weeks ago, Mistral launched Mistral Medium 3, an efficient general-purpose model. Around the same time, the company rolled out Le Chat Enterprise, a corporate-focused chatbot service that offers tools like an AI “agent” builder and integrates Mistral’s models with third-party services like Gmail, Google Drive, and SharePoint.

Microsoft have just made a major announcement that will shake the business model of popular AI based code editors Cursor and Windsurf. Both of these companies are built around forks of Visual Studio Code that build AI tooling “into the bones”. Microsoft have just announced they are open sourcing GitHub Copilot Chat Extension and will be building the functionality into the core of Visual Studio Code. Details from the Visual Studio Code announcement: We believe that the future of code editors should be open and powered by AI. For the last decade, VS Code has been one of the most successful OSS projects on GitHub. We are grateful for our vibrant community of contributors and users who choose VS Code because it is open source. As AI becomes core to the developer experience in VS Code, we intend to stay true to our founding development principles: open, collaborative, and community-driven. We will open source the code in the GitHub Copilot Chat extension under the MIT license, then carefully refactor the relevant components of the extension into VS Code core. This is the next and logical step for us in making VS Code an open source AI editor. It’s a reflection that AI-powered tools are core to how we write code; a reaffirmation of our belief that working in the open leads to a better product for our users and fosters a diverse ecosystem of extensions. Why open source now? Over the last few months, we’ve observed shifts in AI development that motivated us to transition our AI development in VS Code from closed to open source: Large language models have significantly improved, mitigating the need for “secret sauce” prompting strategies. The most popular and effective UX treatments for AI interactions are now common across editors. We want to enable the community to refine and build on these common UI elements by making them available in a stable, open codebase. An ecosystem of open source AI tools and VS Code extensions has emerged. We want to make it easier for these extension authors to build, debug, and test their extensions. This is especially challenging today without access to the source code in the Copilot Chat extension. We’ve gotten a lot of questions about the data that is collected by AI editors. Open sourcing the Copilot Chat extension enables you to see the data we collect, increasing transparency. Malicious actors are increasingly targeting AI developer tools. Throughout VS Code’s history as OSS, community issues and PRs have helped us find and fix security issues quickly. Next steps In the coming weeks, we will work to open source the code in the GitHub Copilot Chat extension and refactor AI features from the extension into VS Code core. Our core priorities remain intact: delivering great performance, powerful extensibility, and an intuitive, beautiful user interface. Open source works best when communities build around a stable, shared foundation. Thus, our goal is to make contributing AI features as simple as contributing to any part of VS Code. The stochastic nature of large language models makes it especially challenging to test AI features and prompt changes. To ease this, we will also make our prompt test infrastructure open source to ensure that community PRs can build and pass tests. In unrelated news, Microsoft also announced the open-sourcing of WSL or Windows Subsystem for Linux. You can learn more in the video below.

Fine-tuning LLMs often requires extensive resources, time, and memory, challenges that can hinder rapid experimentation and deployment. Unsloth AI revolutionizes this process by enabling fast, efficient fine-tuning state-of-the-art models like Qwen3-14B with minimal GPU memory, leveraging advanced techniques such as 4-bit quantization and LoRA (Low-Rank Adaptation). In this tutorial, we walk through a practical implementation on Google Colab to fine-tune Qwen3-14B using a combination of reasoning and instruction-following datasets, combining Unsloth’s FastLanguageModel utilities with trl.SFTTrainer users can achieve powerful fine-tuning performance with just consumer-grade hardware. %%capture import os if "COLAB_" not in "".join(os.environ.keys()): !pip install unsloth else: !pip install --no-deps bitsandbytes accelerate xformers==0.0.29.post3 peft trl==0.15.2 triton cut_cross_entropy unsloth_zoo !pip install sentencepiece protobuf "datasets>=3.4.1" huggingface_hub hf_transfer !pip install --no-deps unsloth We install all the essential libraries required for fine-tuning the Qwen3 model using Unsloth AI. It conditionally installs dependencies based on the environment, using a lightweight approach on Colab to ensure compatibility and reduce overhead. Key components like bitsandbytes, trl, xformers, and unsloth_zoo are included to enable 4-bit quantized training and LoRA-based optimization. from unsloth import FastLanguageModel import torch model, tokenizer = FastLanguageModel.from_pretrained( model_name = "unsloth/Qwen3-14B", max_seq_length = 2048, load_in_4bit = True, load_in_8bit = False, full_finetuning = False, ) We load the Qwen3-14B model using FastLanguageModel from the Unsloth library, which is optimized for efficient fine-tuning. It initializes the model with a context length of 2048 tokens and loads it in 4-bit precision, significantly reducing memory usage. Full fine-tuning is disabled, making it suitable for lightweight parameter-efficient techniques like LoRA. model = FastLanguageModel.get_peft_model( model, r = 32, target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"], lora_alpha = 32, lora_dropout = 0, bias = "none", use_gradient_checkpointing = "unsloth", random_state = 3407, use_rslora = False, loftq_config = None, ) We apply LoRA (Low-Rank Adaptation) to the Qwen3 model using FastLanguageModel.get_peft_model. It injects trainable adapters into specific transformer layers (like q_proj, v_proj, etc.) with a rank of 32, enabling efficient fine-tuning while keeping most model weights frozen. Using “unsloth” gradient checkpointing further optimizes memory usage, making it suitable for training large models on limited hardware. from datasets import load_dataset reasoning_dataset = load_dataset("unsloth/OpenMathReasoning-mini", split="cot") non_reasoning_dataset = load_dataset("mlabonne/FineTome-100k", split="train") We load two pre-curated datasets from the Hugging Face Hub using the library. The reasoning_dataset contains chain-of-thought (CoT) problems from Unsloth’s OpenMathReasoning-mini, designed to enhance logical reasoning in the model. The non_reasoning_dataset pulls general instruction-following data from mlabonne’s FineTome-100k, which helps the model learn broader conversational and task-oriented skills. Together, these datasets support a well-rounded fine-tuning objective. def generate_conversation(examples): problems = examples["problem"] solutions = examples["generated_solution"] conversations = [] for problem, solution in zip(problems, solutions): conversations.append([ {"role": "user", "content": problem}, {"role": "assistant", "content": solution}, ]) return {"conversations": conversations} This function, generate_conversation, transforms raw question–answer pairs from the reasoning dataset into a chat-style format suitable for fine-tuning. For each problem and its corresponding generated solution, a conversation is conducted in which the user asks a question and the assistant provides the answer. The output is a list of dictionaries following the structure expected by chat-based language models, preparing the data for tokenization with a chat template. reasoning_conversations = tokenizer.apply_chat_template( reasoning_dataset["conversations"], tokenize=False, ) from unsloth.chat_templates import standardize_sharegpt dataset = standardize_sharegpt(non_reasoning_dataset) non_reasoning_conversations = tokenizer.apply_chat_template( dataset["conversations"], tokenize=False, ) import pandas as pd chat_percentage = 0.75 non_reasoning_subset = pd.Series(non_reasoning_conversations).sample( int(len(reasoning_conversations) * (1.0 - chat_percentage)), random_state=2407, ) data = pd.concat([ pd.Series(reasoning_conversations), pd.Series(non_reasoning_subset) ]) data.name = "text" We prepare the fine-tuning dataset by converting the reasoning and instruction datasets into a consistent chat format and then combining them. It first applies the tokenizer’s apply_chat_template to convert structured conversations into tokenizable strings. The standardize_sharegpt function normalizes the instruction dataset into a compatible structure. Then, a 75-25 mix is created by sampling 25% of the non-reasoning (instruction) conversations and combining them with the reasoning data. This blend ensures the model is exposed to logical reasoning and general instruction-following tasks, improving its versatility during training. The final combined data is stored as a single-column Pandas Series named “text”. from datasets import Dataset combined_dataset = Dataset.from_pandas(pd.DataFrame(data)) combined_dataset = combined_dataset.shuffle(seed=3407) from trl import SFTTrainer, SFTConfig trainer = SFTTrainer( model=model, tokenizer=tokenizer, train_dataset=combined_dataset, eval_dataset=None, args=SFTConfig( dataset_text_field="text", per_device_train_batch_size=2, gradient_accumulation_steps=4, warmup_steps=5, max_steps=30, learning_rate=2e-4, logging_steps=1, optim="adamw_8bit", weight_decay=0.01, lr_scheduler_type="linear", seed=3407, report_to="none", ) ) We take the preprocessed conversations, wrap them into a Hugging Face Dataset (ensuring the data is in a consistent format), and shuffle the dataset with a fixed seed for reproducibility. Then, the fine-tuning trainer is initialized using trl’s SFTTrainer and SFTConfig. The trainer is set up to use the combined dataset (with the text column field named “text”) and defines training hyperparameters like batch size, gradient accumulation, number of warmup and training steps, learning rate, optimizer parameters, and a linear learning rate scheduler. This configuration is geared towards efficient fine-tuning while maintaining reproducibility and logging minimal details (with report_to=”none”). trainer.train() trainer.train() starts the fine-tuning process for the Qwen3-14B model using the SFTTrainer. It trains the model on the prepared mixed dataset of reasoning and instruction-following conversations, optimizing only the LoRA-adapted parameters thanks to the underlying Unsloth setup. Training will proceed according to the configuration specified earlier (e.g., max_steps=30, batch_size=2, lr=2e-4), and progress will be printed every logging step. This final command launches the actual model adaptation based on your custom data. model.save_pretrained("qwen3-finetuned-colab") tokenizer.save_pretrained("qwen3-finetuned-colab") We save the fine-tuned model and tokenizer locally to the “qwen3-finetuned-colab” directory. By calling save_pretrained(), the adapted weights and tokenizer configuration can be reloaded later for inference or further training, locally or for uploading to the Hugging Face Hub. In conclusion, with the help of Unsloth AI, fine-tuning massive LLMs like Qwen3-14B becomes feasible, using limited resources, and is highly efficient and accessible. This tutorial demonstrated how to load a 4-bit quantized version of the model, apply structured chat templates, mix multiple datasets for better generalization, and train using TRL’s SFTTrainer. Whether you’re building custom assistants or specialized domain models, Unsloth’s tools dramatically reduce the barrier to fine-tuning at scale. As open-source fine-tuning ecosystems evolve, Unsloth continues to lead the way in making LLM training faster, cheaper, and more practical for everyone. Check out the COLAB NOTEBOOK. 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/Chain-of-Thought May Not Be a Window into AI’s Reasoning: Anthropic’s New Study Reveals Hidden GapsAsif Razzaqhttps://www.marktechpost.com/author/6flvq/How to Build a Powerful and Intelligent Question-Answering System by Using Tavily Search API, Chroma, Google Gemini LLMs, and the LangChain FrameworkAsif Razzaqhttps://www.marktechpost.com/author/6flvq/AWS Open-Sources Strands Agents SDK to Simplify AI Agent DevelopmentAsif Razzaqhttps://www.marktechpost.com/author/6flvq/Windsurf Launches SWE-1: A Frontier AI Model Family for End-to-End Software Engineering 🚨 Build GenAI you can trust. ⭐️ Parlant is your open-source engine for controlled, compliant, and purposeful AI conversations — Star Parlant on GitHub! (Promoted)