The coding model an AI mostly built itself

Flint is a mixture-of-experts model built for agentic coding on local hardware — investigating codebases, fixing bugs, implementing features, writing tests. The full model is 40 billion parameters, but only about 2.5 billion are active for any given token, which is how it stays fast enough for one machine. On context, the sources differ: Recursive's launch post says Flint supports windows of up to 512k tokens and recommends staying within 256k, while the Hugging Face model card documents a 262,144-token context configuration, prompt and output included. Neither explains the gap.

The weights are on Hugging Face under the Apache 2.0 license, so anyone can download, modify, and ship them commercially. It runs through llama.cpp, Ollama, Unsloth, LM Studio, vLLM, and SGLang, and Recursive is shipping its own lightweight inference engine, rec, tuned for Flint on M4 Pro and M5 Pro/Max chips. Apple Silicon users need 48GB of unified memory; on the NVIDIA side, 48GB GPUs.

Recursive is the AI startup co-founded last year by Richard Socher, Salesforce's former chief scientist. It emerged from stealth in May with $650 million at a $4.65 billion valuation from GV, Nvidia, AMD Ventures, and Greycroft. The company's whole bet is in its name: AI systems that improve themselves.

Speed first, since that's the selling point: with Recursive's inference engine, Flint generates more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 per second on a 64GB NVIDIA RTX Spark, at short contexts. The company says that's up to twice as fast as comparable models and more than five times faster than dense 25-to-40-billion-parameter models with similar coding scores.

On benchmarks — all company-reported, none independently verified yet — Recursive claims Flint leads every mixture-of-experts model under 50 billion parameters on agentic coding: 74.2% on SWE-Bench Verified, 89.7% on SWE-bench Pro v2, 54.9% on Terminal-Bench 2.1, 95.1% on IFEval. Dense models still win some matchups; Alibaba's Qwen 3.8 outscores it on Terminal-Bench 2.1.

Recursive is upfront about the limitation: "While Flint is very capable, it may need more guidance than the larger closed-source coding models you're used to." Their advice is to break complex work into smaller subtasks with explicit instructions. Independent signal is thin so far, though Wired's Will Knight built several small programs on a 48GB Mac and wrote that "Flint nailed my coding requests every time."

This is the part worth reading closely. The Eureka Machine automates the full research loop: proposing ideas, implementing them, testing them, and using the results to decide what to try next. As Socher told Wired, "Nearly every aspect of the entire model pipeline was actually done and improved upon by AI." Recursive credits it across the whole stack.

Three concrete examples: on architecture, it searched expert layouts and widths, moving the design from a 35B-total/3B-active baseline to 40B total with only 2.5B active — more capacity, less work per token. On quantization, it pushed even the output layer down to 4 bits (where higher precision is usually kept), then recovered accuracy with an offline distillation stage that nearly halved the training steps compared with its quantization-aware training baseline — the final NVFP4 model scores within 0.7 points of the full 16-bit model. On inference, it wrote custom kernels, including a Metal megakernel that fuses the entire decode pass on M4/M5 Macs. That kernel code is public too.

The timeline is the other number that matters: a small team of people and agents built Flint from scratch in under four months. That is fast for a from-scratch model, and it's the entire thesis of the company — better models make the Eureka Machine a better researcher, which builds the next model faster.

Flint is a modest first test of recursive self-improvement, not the full loop — humans oversaw the process and Socher says the long-term goal keeps humans able to stop it. But it's a public, downloadable artifact of an AI research system that, Recursive says, contributed to architecture, training, and inference work. The real test isn't the launch charts; it's what developers find when they run it on their own laptops and compare it against models built the old-fashioned way.

The practical pitch is simpler, and it's in Socher's launch post on X: "Private, unmetered intelligence on your own machine." Code stays local, no API meter runs. If you're curious, there's a free trial on OpenCode and pi.dev, and the open weights mean the community can now stress-test the Eureka Machine's work directly.