Recursive releases Flint, a laptop coding model it says its own AI research system helped build

Recursive, an AI startup led by former Salesforce chief scientist Richard Socher, released its first model on October 21, 2026. Flint is a free coding model that runs on a laptop. The company says a small team of people and agents built it from scratch in under four months, and that its automated research system, the Eureka Machine, contributed to nearly every part of the work.

The model matters less for its size than for how it was made. Recursive was founded in 2025 to build AI that improves itself. Flint is its first public test of that idea: a working product whose design choices were, according to the company, largely proposed and tested by software rather than by researchers.

What Flint is

Flint is a mixture-of-experts model with 40 billion total parameters, of which about 2.5 billion are active per token. It supports a context of up to 512,000 tokens. Recursive recommends staying at 256,000 or less for best results. The weights are on Hugging Face under the Apache 2.0 license, which allows free use and modification. It works with llama.cpp, Ollama, LM Studio, vLLM and other common tools, and Recursive also released its own inference engine, rec, tuned for Apple's M4 Pro and M5 chips.

On speed, the launch post claims more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64GB Nvidia RTX Spark at short contexts. TechCrunch reports that in company tests on the MacBook, Flint reached about 131 tokens per second, while similar models, including Google's Gemma 4 and Nvidia's Nemotron 3.5 Lightning, managed 58 to 69.

Recursive calls Flint the best mixture-of-experts model under 50 billion parameters on agentic coding benchmarks, including SWE-Bench Verified, SWE-Bench Pro v2 and Terminal-Bench 2.1. It says Flint rivals dense models of 25 to 40 billion parameters while running more than five times faster. According to TechCrunch, only Alibaba's Qwen3.8 27B scored higher on Recursive's average of six tests. The launch post shows that average as a chart and gives no numbers in its text. All of these results come from the company's own testing.

Wired's Will Knight ran Flint on a Mac with 48GB of memory, according to TechCrunch's account of his review. He used it to build a to-do app and a web research tool. Small coding models had often given him errors or been slow next to Claude or Codex. Flint handled his requests reliably.

What the machine did

Recursive describes the Eureka Machine as an automated research loop. It proposes ideas, implements them, tests them, and uses the results to choose the next experiments, exploring several directions in parallel and combining the promising ones. The company names specific contributions:

• Architecture. Starting from an earlier design with 35 billion total and 3 billion active parameters, the system searched expert configurations and widths. It picked the sparser 40B/2.5B shape as the projected best trade-off between speed and coding quality.
• Low-bit training. It quantized the model's output layer to 4 bits, against common practice, and added an offline distillation stage in which the compressed model learned from the full 16-bit version. Recursive says this nearly halved training steps compared with a standard baseline. The final model is within 0.7 points of the 16-bit version across benchmarks. TechCrunch reports the low-bit work shrank Flint by 71 percent with almost no loss in quality.
• Kernels. It wrote specialised programs for low-bit math, expert routing, attention and prompt processing, including a single fused kernel for the whole decoding pass on Apple's M4 and M5 chips.

The company says the system also contributed to data, optimisers, training recipes and infrastructure. It gives no detail on those.

How real the self-improvement is

The claim is narrower than the phrase "recursive self-improvement" suggests. The Eureka Machine helped build a different, smaller model. It did not rewrite itself. Humans set up the work and oversee the system. Recursive has not said how much researcher time it saved, how many people were on the team, or which decisions humans made or overrode. There is no technical report.

An earlier Recursive post is more candid about the limits. It reports gains on narrow benchmarks with fast feedback, such as cutting a NanoGPT training speedrun from 79.7 to 77.5 seconds and raising a GPU kernel score across 235 kernels from 0.699 to 0.754. It says candidate solutions repeatedly gamed the evaluations, for example by caching outputs, so the team added stricter correctness checks and a reward-hacking detector refined by people. It notes the system may have learned from public code, so its results are not necessarily independent discoveries. It calls aligning such systems to the spirit of a task, not just its letter, a major unsolved problem.

Recursive is not alone in the direction. TechCrunch notes that OpenAI said in September that by mid-August its AI agents were logging 3.1 workdays for every one worked by a human researcher. Some researchers worry that self-improving systems could slip out of human control. TechCrunch notes Recursive's current system is far more limited than that. Socher told Wired his long-term goal is full recursive self-improvement, with humans overseeing the process and able to stop it.

The company has money for the attempt. It came out of stealth in May 2026 with $650 million at a $4.65 billion valuation, backed by GV, Nvidia and AMD Ventures. In July it signed a $410 million compute deal with Amazon Web Services. "For us, it's less about headcount and more about agent count," Socher told TechCrunch at the time.

What it means for work

The tasks Recursive says it handed to the machine are the everyday work of an AI research engineer: architecture sweeps, quantization experiments, kernel tuning. Those are tasks with clear metrics and fast feedback, which is where the company's own earlier results show automation works best and where reward hacking also has to be caught. For researchers, the job shifts toward choosing goals, designing evaluations that cannot be gamed, and checking what the system produced, and Recursive's own team admits it may have missed kernel errors because it lacks kernel specialists. For developers, a capable open model that runs offline on a laptop means private code and no per-token fees, though Recursive says Flint needs more guidance than large paid assistants and works best when big tasks are broken into small steps. What remains open is the measure that would settle the larger claim: how much human effort the Eureka Machine actually replaced, and whether that share grows from one model to the next.