Recursive Flint: The 40B Open Coding Model 'Built by AI' That Runs on a MacBook

Recursive, Richard Socher's startup, released Flint on October 21, 2026: an Apache 2.0, open-weight coding model with 39.8B total and about 2.5B active parameters. Recursive says its automated research system, the Eureka Machine, did much of the work. The company reports 89.7% on SWE-bench Pro v2 and over 125 tokens per second on a MacBook Pro, but all figures are self-reported.

What is Recursive Flint?

Flint is a mixture-of-experts language model built for coding agents, the kind that read codebases, fix bugs, implement features and write tests. The Hugging Face model card gives these specs:

• Size: 39.8B total parameters and about 2.5B active per token, rounded to "40B-A2.5B."
• Architecture: 49 layers, with 256 routed experts per MoE layer, 6 picked per token, plus one shared expert.
• Context: 262,144 tokens shared between prompt and output.
• Training: trained from scratch, with a stated knowledge cutoff of June 2026.
• License: Apache 2.0 for the weights.

Flint is Recursive's first model. Crypto Briefing reports the company has raised $650 million at a $4.65 billion valuation, with backers including GV, Greycroft, Nvidia and AMD Ventures.

The sources disagree on context length. Recursive's launch post lists a 512k maximum and recommends staying within 256k, while the model card says 262,144. Plan around 256k.

How well does Flint perform, and can the numbers be trusted?

Flint's reported scores are strong, but Recursive measured them itself, and no independent evaluation had been published as of October 23. The model card lists:

• SWE-bench Pro v2: 89.7%
• SWE-bench Verified: 74.2%
• Terminal-Bench 2.1: 54.9%

These are average pass@1 scores over at least five attempts, with agent timeouts of up to five hours. SWE-bench ran without internet access. On Terminal-Bench, solutions found through web retrieval were counted as failures. The model card warns that other developers may not apply the same adjustments, so cross-model comparisons are not like-for-like.

There are reasons to wait for outside testing:

1. The score on Pro v2 is higher than the score on Verified, even though Verified is usually seen as the easier set. Differences between benchmark versions may explain this, but reproduction would settle it.
2. The launch post names no rival models in its text. It says only that Flint "leads MoE models with fewer than 50 billion total parameters."
3. RuntimeWire notes the company has not independently verified its performance and speed claims.

What hardware do you need to run Flint locally?

For agentic coding, plan on about 48GB of memory. Smaller quantized builds will load with less:

• Recursive's rec engine: the rec repo requires an M4 Pro, M5 Pro or M5 Max Mac with at least 48GB of unified memory and about 30GB of disk.
• LM Studio: the LM Studio listing gives a 24GB minimum for its GGUF build.
• Smallest build: AI Weekly reports quantizations as small as 11.4GB.
• Full precision: about an 80GB download, served on one B300 or H200 GPU.

Recursive's speed claims (over 125 tok/s on an M4 Pro MacBook Pro, over 150 on a 64GB RTX Spark at short contexts) exclude prefill, the time spent reading the prompt. Expect longer waits on big prompts.

How do you run Flint?

You can run Flint in four main ways:

1. rec on a Mac: install it, then run "rec launch pi" or "rec launch opencode". The first run downloads the model and the coding harness.
2. LM Studio, Ollama or llama.cpp: use the community GGUF quantizations.
3. vLLM or SGLang: serve the checkpoint through an OpenAI-compatible endpoint.
4. Docker Model Runner: pull the model straight from Hugging Face.

Recommended settings are temperature 1.0, top-p 0.95, top-k 20, and an output budget of at least 32,768 tokens for hard problems.

How was Flint "built by AI"?

Recursive says the Eureka Machine automated the research loop: proposing ideas, implementing them, testing them and choosing what to try next. According to the launch post, its work included:

• Architecture search: it tested expert counts and widths, starting from a 35B-total, 3B-active baseline.
• Quantization: it quantized the output layer to 4 bits and added a distillation stage. Recursive says the final NVFP4 build lands within 0.7 points of the 16-bit model.
• Kernels: it wrote a Metal megakernel for M4 and M5 chips that fuses the whole decode pass.

Recursive credits a "small team of people and agents" and says Flint was built in under four months. It gives no team size and does not say which decisions people made, and the Hugging Face model card does not mention the Eureka Machine at all. The fair reading is that experiments and engineering were heavily automated under human oversight, not that the model designed itself.

What are Flint's limitations?

Recursive itself says Flint may need more guidance than larger closed coding models. Other caveats:

• The company recommends splitting complex work into smaller subtasks with explicit instructions.
• Long reasoning can add latency and use up the output budget before an answer is finished.
• A long context window does not guarantee equal accuracy across the whole window.
• Tool-using apps need their own permissions and sandboxing.
• Results vary with precision and runtime, so test the exact checkpoint you plan to deploy.

What does Flint mean for marketing teams building with AI?

Models like Flint lower two barriers to in-house automation: per-token cost and data exposure. When a capable model runs on one laptop with no API key, scripts and AI agents in marketing can work on first-party data that never leaves the machine.

The model is rarely what decides success. Teams that get value from agents have clean data, clearly defined tasks, review steps and a way to measure results, and a few benchmark points matter less than whether your workflow catches bad output. Our AI workforce guide covers how to set that up.

More open models also means more places where answers about your brand get generated, so it gets more useful to track which models recommend you. That is what model share measures. For a closed-model comparison, see our Gemini 4 Argon breakdown.

FAQ

Is Recursive Flint free to use?

Yes. The weights are released under Apache 2.0, which allows commercial use. Running Flint locally costs only hardware and electricity.

How many parameters does Flint have?

Flint has 39.8B total parameters and activates about 2.5B per token through its mixture-of-experts design.

Can Flint run on a MacBook?

Yes, on recent high-memory models. Recursive's rec engine needs an M4 Pro, M5 Pro or M5 Max with at least 48GB of unified memory.

Is the 89.7% SWE-bench Pro v2 score verified?

No. Recursive measured and reported the score itself, and no independent evaluation had been published as of October 23, 2026.

What is the Eureka Machine?

The Eureka Machine is Recursive's automated AI research system. The company says it proposes, implements and tests ideas, and that it contributed to Flint's architecture, quantization and kernels.

What context length does Flint support?

The model card lists 262,144 tokens. Recursive's launch post mentions up to 512k but recommends staying within 256k.

Last updated October 23, 2026.