AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Recursive launches Flint, a laptop coding model it says was largely built by AI

Collected Oct 21, 2026

Recursive, an AI startup founded last year, released its first model on Wednesday. Called Flint, it is a free coding model designed to run on a laptop, and the company says it was built largely with the help of its own AI. Recursive's automated research system, which it calls the Eureka Machine, contributed to nearly every step of development, allowing a small team to build the model from scratch in under four months. The release is positioned as an early public test of recursive self-improvement, the idea that AI can build better AI that in turn builds better AI.

The model is free to download on Hugging Face under the Apache 2.0 license, which permits free use and modification. Because Flint runs on a user's own machine rather than a company's cloud servers, developers can work offline, keep code private, and avoid usage fees that cloud AI services charge based on how much text the model processes. Recursive emerged from stealth in May with $650 million in funding at a $4.65 billion valuation from backers including GV, Nvidia and AMD Ventures, and is led by CEO Richard Socher, Salesforce's former chief scientist.

Architecturally, Flint is a mixture-of-experts model. That means it activates only a small slice of itself for any given task: 2.5 billion of its 40 billion parameters at a time. The approach is common in large open models, where a router sends each token to a subset of expert sub-networks, so total parameter count sets memory needs while active parameters govern compute per token. In practice this is what makes Flint light enough for an ordinary laptop and far smaller than Meta's Muse Spark Open, an open model built for Macs released last week, which uses 816 billion parameters and needs at least a terabyte of memory.

Recursive also reports speed advantages. In its tests, Flint generated about 131 tokens per second on a MacBook Pro with Apple's M4 Pro chip, roughly twice the 58 to 69 tokens per second that similar models managed on the same machine, including Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning. On Nvidia's RTX Spark, Flint topped 150 tokens per second. Token throughput is a useful but imperfect proxy for a coding assistant's real-world feel, since latency, tool-calling behavior and how often a model needs to regenerate a failed edit matter as much as raw speed.

On benchmarks, Recursive says Flint outperformed other models under 50 billion parameters on SWE-Bench Verified, which measures how well AI can fix real-world software bugs. On the company's own average across six tests, only one model, Alibaba's Qwen3.8 27B, scored higher. Company-run benchmark suites deserve some skepticism because the selection of tasks and scoring can favor the model under test, but the SWE-Bench result is at least a comparable, widely used measure.

Recursive acknowledged that Flint needs more guidance than the large paid coding assistants and suggested breaking large tasks into smaller steps. Early reviews were positive: Wired's Will Knight used Flint on a Mac with 48GB of memory to build several small programs, including a to-do list app and a tool that scours the web for research, and found that while smaller coding models often ran into errors or were slow compared with Claude or Codex, Flint handled his coding requests every time.

The Eureka Machine itself comes up with ideas, runs experiments and uses the results to decide what to try next. According to the company's launch video, it contributed to nearly every part of building Flint, spanning pre-training, post-training and inference. Recursive credits it with specific improvements: a sparser model shape, low-bit training that shrank Flint by 71% with almost no loss in quality, and custom kernels, which are specialized programs that speed up the underlying computations. Recursive did not say how much time Eureka saved its researchers.

The company is not alone in leaning on AI for research work. OpenAI said last month that by mid-August its AI agents were logging 3.1 workdays for every one worked by a human researcher. Recursive previously told TechCrunch that its $410 million compute deal with Amazon Web Services in June was less about headcount and more about agent count.

The longer-term goal, Socher told Wired, is for the Eureka Machine to reach full recursive self-improvement, with humans overseeing the process and able to stop it if something goes wrong. Recursive says each new model should make the Eureka Machine a better researcher, helping it build the next one faster. That ambition is among the more debated ideas in the industry, with some researchers concerned that self-improving AI could slip out of human control; Recursive's current version is far more modest, an AI research assistant overseen by people that produced a model small enough for a laptop. A likely trade-off for users is that a 40-billion-parameter model with 2.5 billion active parameters will not match frontier cloud assistants on long, multi-file refactors, which is consistent with the company's advice to decompose tasks.

Why it matters: Developers who want a private, offline coding assistant without per-token cloud bills now have an openly licensed option that runs on everyday hardware, which is most relevant for teams working with sensitive code or constrained budgets. The release is also an early, concrete data point for the self-improvement debate: Recursive is claiming real gains from AI-assisted research rather than a purely human build, though the results are self-reported and the model is deliberately small. Whether each generation of Flint meaningfully speeds up Eureka, as the company expects, is the claim to watch in future releases.

Read at TechCrunch · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Startup Recursive has released Flint, a free, laptop-sized coding model built largely by its own AI research system, as an early test of AI that improves itself.