AI coding just hit a milestone nobody’s pricing model was ready for: a competent coding model that runs entirely on a laptop — private, offline, no per-token fees — and that its maker says was largely built by AI itself. Recursive, the startup founded by former Salesforce chief scientist Richard Socher, launched Flint this week: a free, open-weights sparse model (40 billion total parameters, 2.5 billion active) that generated about 131 tokens per second on a MacBook Pro with Apple’s M4 Pro chip — roughly double the 58-69 tokens per second that rivals like Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning managed on the same machine — while outperforming every other model under 50 billion parameters on the industry’s real bug-fixing benchmark. Beneath the benchmarks sits the more consequential claim: Flint’s development was assisted at nearly every step by Recursive’s automated research system, the Eureka Machine, letting a small team build the model from scratch in under four months.
Key Takeaway
- The model: Flint is a free, open-weights sparse mixture-of-experts model — 40B total, 2.5B active parameters — built by Recursive, the startup led by Richard Socher, Salesforce’s former chief scientist.
- The laptop story: about 131 tokens/second on a MacBook Pro M4 Pro (roughly twice Gemma 4’s and Nemotron 3.5 Lightning’s pace on the same machine), over 150 tok/s on Nvidia’s RTX Spark — with offline use, private code, and zero per-token fees.
- The capability: on SWE-Bench Verified — real-world bug-fixing — Flint outperformed every comparable model under 50 billion parameters; on Recursive’s six-test average only Alibaba’s Qwen3.8 27B scored higher.
- The origin claim: the Eureka Machine, Recursive’s automated research system, helped with nearly every step of development — a small team shipped the model in under four months, an early public test of AI building AI.
- The money: $650 million raised at a $4.65 billion valuation from GV, Nvidia and AMD Ventures — a strong signal that hardware-side players see local AI coding as a market, not a demo.
What Flint actually is
Strip the launch-day language and Flint is a deliberately engineered trade-off: a sparse mixture-of-experts architecture that keeps the parameter count large enough to be capable (40 billion total) while activating only 2.5 billion per token — small enough, in practice, to run on consumer hardware instead of a rented GPU. That single design choice cascades into everything users care about: the model runs on your computer, so your code never leaves your machine; you can work offline; and you pay nothing per token, because the compute is yours. In an AI coding market dominated by metered cloud subscriptions, Flint’s pitch is ownership rather than rental — a model you possess instead of a service you stream.
The market math explains the funding. AI coding assistants have become the fastest-adopted developer tools in history, and every mainstream option bills by usage — a cost model that scales with your ambition. A capable local model converts that recurring bill into a one-time hardware expense, which is why Flint’s release matters to budget-sensitive developers even before anyone debates its benchmark standing against the cloud giants.
The numbers that matter: speed, privacy, price
Speed first, because latency is where local models usually lose. In Recursive’s own tests, Flint generated about 131 tokens per second on a MacBook Pro with an M4 Pro chip — roughly twice as fast as Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning on identical hardware (58-69 tok/s) — and topped 150 tokens per second on Nvidia’s RTX Spark. Interactive coding lives or dies on that responsiveness: an assistant that answers before your coffee cools changes how a workday feels.
Capability second. On SWE-Bench Verified — the benchmark that scores how well an AI fixes real-world software bugs — Flint outperformed every comparable model under 50 billion parameters. On Recursive’s own six-test average, a single model scored higher: Alibaba’s Qwen3.8 27B. That is a respectable, honest result — the best-in-class small model, explicitly not beating the cloud-scale frontier. Recursive acknowledges Flint was produced with substantial human oversight, so the AI-built-AI story is assistance at scale, not autonomy; the company’s framing is careful and the distinction matters for how the claim gets used.
Price third, and here Flint is categorical: the model is free to download and run, with open weights, and no metered backend exists to bill you. The privacy corollary is the part enterprise buyers will actually read: code, context and credentials stay on-device — a compliance posture cloud services cannot match at any price.
The claim of AI-built AI, examined
The Eureka Machine is Recursive’s automated research system: pipelines that handle most steps of model development — data curation, training runs, evaluation, iteration — with humans supervising rather than hand-driving. The company says it helped with nearly every step of Flint’s creation, compressing what is normally a year-plus effort into under four months with a small team. Recursive frames this explicitly as an early public test of recursive self-improvement: AI building better AI, building better AI (TechCrunch’s launch report carries the full benchmark tables). The claim’s honest reading sits between the hype and the dismissal. What is genuinely new is the cycle-time compression — a small lab shipping a frontier-adjacent artifact quarterly rather than annually. What is not yet demonstrated is a closed loop with no human in it; Wired’s coverage, and Socher’s own positioning, treat the distinction as the story’s live question rather than a settled fact — and Wired’s feature makes the case for taking it seriously either way.
Why it matters beyond one startup: if the Eureka-style pipeline holds, small teams everywhere can iterate model releases at software-iteration speed — the moat shifts from scale of capital to quality of automation. The practical consequence for the AI coding market specifically is supply: more capable small models, faster, from more labs — which compresses prices everywhere, and not just for the open-weight community. The cloud providers’ metered businesses now face a floor they cannot bill below.
The investors and the momentum
Recursive emerged from stealth in May with $650 million at a $4.65 billion valuation, from GV (Google Ventures), Nvidia and AMD Ventures. Read that list vertically: Google Ventures funds the software bet, while Nvidia and AMD — the two companies that supply the silicon local models run on — bankroll a future where AI coding happens on their consumer and workstation hardware rather than exclusively in hyperscaler clouds. Socher’s own pedigree (Salesforce’s former chief scientist) gives the team research credibility from day one. The strategic shape is coherent: chips makers need reasons for consumers to buy premium silicon; local AI coding is one of the strongest reasons emerging; Flint is simultaneously a product and an argument for the hardware beneath it.
What it means for developers and freelancers
For working developers, the decision framework is simpler than the discourse: Flint is worth testing the moment your subscription bill exceeds your hardware’s amortized cost — and worth adopting for privacy-bound work regardless, since local execution is the only zero-leak architecture. For freelancers whose clients include regulated industries (health, finance, legal) and whose work must respect the open-weight licensing boundaries our AI model licenses guide explains, AI coding that never phones home changes what you can offer under NDA. For teams training the next generation, it changes the economics of teaching — the same logic our AI-certifications guide applies to learning: the tools that run where you are, cheaper, win the adoption race.
And for the broader earn-with-AI crowd our get-paid-to-train-AI coverage serves: smaller models needing efficient reinforcement from human reviewers keep that labor market growing even as inference migrates to devices. The chips beneath it all trace back through the supply chain our EUV lithography guide mapped — every local-model milestone is, upstream, a fab-schedule story. AI coding on the desk, AI-built pipelines in the lab, and the silicon queue behind both: the three layers of this week’s release are one story told at three depths.
How to try Flint this week
The test path is short. Check the hardware floor first: an M-series Mac with at least the memory footprint the model needs on your machine, or a Windows/Linux box with a modern RTX-class GPU — the RTX Spark path is the one Recursive benchmarks on the PC side. Download the open weights, run the model locally, and point it at a real workday: a refactoring task you have been avoiding, a legacy file only your team understands, a test suite with gaps. Time it against your current metered assistant on the same tasks: if the tokens-per-second hold and the quality lands close, the subscription math changes that week. Keep two honest checks in the loop — compare outputs on your hardest real ticket rather than toy prompts, and read what the model refuses or fumbles, because that boundary is exactly the line between laptop-class and cloud-class models today.
For teams, the pilot design is the one our freelance-AI coverage uses for any tooling shift: one engineer, two weeks, three measurable tickets, a written verdict. Local AI coding earns its rollout the same way any tool earns it — by beating the incumbent on your actual workload, not on launch-day numbers. And watch the ecosystem around the weights: community fine-tunes, quantized variants, and editor integrations historically arrive within weeks of an open-weights release, which is when the practical story usually gets good.
The strategic frame for buyers: local AI coding is not a replacement decision, it is a portfolio decision — the metered cloud assistant stays for frontier-hard work; the local model takes the volume tasks, the privacy-bound tasks, and the offline ones. The teams that publish their internal benchmarks this quarter will set the defaults everyone else inherits, and the vendors watching those benchmarks are the same ones repricing subscriptions. AI coding just became a two-column spreadsheet: cloud and local, each with its price column, each winning a different half of your week.
FAQ
What is Recursive’s Flint?
Flint is a free, open-weights AI coding model — a 40-billion-parameter sparse mixture-of-experts model with 2.5 billion active parameters — that runs on consumer hardware like a MacBook Pro M4 Pro, built by the startup Recursive, led by former Salesforce chief scientist Richard Socher.
Can an AI coding model really run on a laptop?
Yes — Flint generated about 131 tokens per second on a MacBook Pro with an M4 Pro chip in Recursive’s tests, roughly double the speed of Google’s Gemma 4 and Nvidia’s Nemotron 3.5 Lightning on the same machine, and topped 150 tokens per second on Nvidia’s RTX Spark hardware.
Was Flint really built by AI?
Largely assisted, with humans supervising: Recursive says its Eureka Machine automated research system helped with nearly every step of development, letting a small team build the model in under four months. The company frames it as an early test of recursive self-improvement — not a closed loop without human oversight.
How good is Flint at coding?
On SWE-Bench Verified, the real-world bug-fixing benchmark, Flint outperformed every comparable model under 50 billion parameters. On Recursive’s six-test average, only Alibaba’s Qwen3.8 27B scored higher. It is the strongest small open-weights model in its class — not a replacement for cloud-scale frontier models.
Does Flint cost anything to use?
No. The model is free to download with open weights, runs on your own hardware, and has no per-token fees — which also means your code stays private and can be used offline, a compliance advantage cloud AI coding services cannot match.
Who funded Recursive?
The company emerged from stealth in May with $650 million at a $4.65 billion valuation, backed by GV (Google Ventures), Nvidia and AMD Ventures — a lineup that pairs the software bet with the chip makers whose hardware local AI coding runs on.
