Recursive Flint: Free 40B Laptop Coding Model Built Largely by AI

The name on the door says it all. Recursive — the AI startup trying to build artificial intelligence that improves itself — released its first model on October 21, 2026. It's called Flint, and it is a free coding model designed to run on a laptop. The part that makes people stop scrolling: Recursive says its own automated research system, called the Eureka Machine, helped with nearly every step of building it, letting a small team build a frontier-class coding model from scratch in under four months.

This is the first public test of recursive self-improvement — AI helping to build better AI — shipped as a download you can run offline on your own machine. I went through the launch post, the model card, and the independent reporting to separate what's proven from what's marketing.

The quick answer

• What launched: Flint, a mixture-of-experts coding model with 40 billion total parameters, of which only about 2.5 billion are active per token — which is why it runs on a laptop.
• What it costs: nothing. The weights are free on Hugging Face under the Apache 2.0 license. No API fees, no per-token billing. The real cost is hardware: you need a 48 GB Mac or a 48 GB NVIDIA GPU.
• How fast: Recursive reports 125+ tokens per second on an M4 Pro MacBook Pro and 150+ tokens per second on a 64GB NVIDIA RTX Spark at short contexts — roughly twice the speed of comparable models.
• How good: Company-reported 74.2% on SWE-bench Verified and 89.7% on SWE-bench Pro v2, leading MoE models under 50B parameters on agentic coding — but these are the company's own numbers, not independently verified.
• The catch: Recursive itself admits Flint needs more hand-holding than the big paid coding assistants, and the benchmark claims haven't been checked by anyone outside the company.

What Flint actually is

Flint is a decoder-only mixture-of-experts Transformer built for one job: agentic coding. It investigates codebases, fixes bugs, implements features, and writes tests. It is text-only, supports reasoning, and speaks the OpenAI tool-call format, so it slots into the agentic coding tools developers already use.

The mixture-of-experts design is the whole trick. Only a small slice of the model — about 2.5 billion of its 40 billion parameters — activates for each token it produces. That sparsity is what lets a 40-billion-parameter model fit on consumer hardware. Recursive's own framing: Flint rivals the coding performance of dense models with 25–40 billion parameters while running more than five times faster.

The spec sheet, taken from Recursive's launch post:

• 40 billion total parameters (39.8B on the model card), ~2.5B active per token
• Context window up to 512K tokens, with 256K recommended for best results
• 125+ tokens/sec on M4 Pro MacBook Pro; 150+ tokens/sec on 64GB RTX Spark
• 4-bit NVFP4 build that scores within 0.7 points of the full 16-bit model
• Apache 2.0 license — free to use, modify, and ship commercially
• Works with llama.cpp, Ollama, Unsloth, LM Studio, vLLM, and SGLang

The Eureka Machine: the part that matters most

The model is interesting. How it was built is the story. Recursive's Eureka Machine automates the full research loop: proposing ideas, implementing them, testing them, and using the results to decide what to try next. It runs many experiments in parallel and builds on earlier results.

According to the company's launch post, the Eureka Machine contributed at nearly every layer of the stack:

• Architecture. It searched expert configurations and settled on the 40B-total / 2.5B-active design, starting from a 35B/3B baseline — projected to offer the best combination of speed and coding performance.
• Quantization. It found an unusual recipe: quantizing the output layer (lm_head) to 4 bits — against common practice — then recovering accuracy with an offline distillation stage. The NVFP4 build scores within 0.7 points of the 16-bit model.
• Kernels. It wrote custom code tuned for Apple Silicon and NVIDIA RTX Spark chips — specialized kernels for low-bit projections, MoE operations, attention, and prompt processing, including a Metal megakernel that fuses the entire decode forward pass on M4 and M5 chips.
• Training. It worked on data, optimizers, and training recipes.

Recursive's bet is compounding: each new model should make the Eureka Machine a better researcher, which builds the next model faster. The company didn't say how much time the system saved its researchers, but context from elsewhere in the industry suggests the direction is real — OpenAI said last month that by mid-August, its AI agents were logging 3.1 workdays of research for every one worked by a human researcher.

The benchmarks: strong, but read the footnotes

Recursive's own results put Flint at 89.7% on SWE-bench Pro v2, 74.2% on SWE-bench Verified, and 54.9% on Terminal-Bench 2.1, plus 95.1% on IFEval. The company says Flint leads every mixture-of-experts model under 50 billion total parameters on agentic coding — and that across its six-benchmark average, only Alibaba's Qwen3.8 27B scored higher.

Three honesty notes before you treat any of this as fact:

1. These are company-run numbers. The model card itself says comparison scores are developer-reported, and rival figures were measured under different setups. Independent replication hasn't happened yet.
2. The weak spot is visible in the company's own chart. Terminal-Bench 2.1 at 54.9% lags the dense Qwen 3.8's 73.0% — a real gap, not hidden.
3. Speed claims are on company hardware and test setups. Independent tests found roughly 131 tokens/sec on an M4 Pro, which matches Recursive's claim — a good sign, but one data point.

The early hands-on evidence is genuinely encouraging. Wired's Will Knight ran Flint on a Mac with 48GB of memory, building small programs including a to-do app and a research tool. He found smaller coding models usually tripped over errors or ran slowly next to Claude or Codex — but Flint, in his words, nailed his coding requests every time.

The trade-offs, stated plainly

Pros

• Truly free and open. Apache 2.0 weights on Hugging Face — download, run offline, modify, ship commercially. No account, no API key, no per-token bill.
• Private by construction. Your code never leaves your machine — a real advantage over cloud coding assistants for anyone working on proprietary code.
• Fast for its class. About twice the token throughput of comparable local models on the same hardware, per both company tests and early independent runs.
• Tool-compatible. OpenAI-compatible tool calls and broad framework support (Ollama, vLLM, LM Studio, llama.cpp) mean it drops into existing agent setups.
• Serious quantization work. The 4-bit build losing under a point of quality is a genuine engineering result, not a rounding trick.

Cons

• Demanding hardware. "Free" assumes you own a 48GB Apple Silicon Mac or a 48GB NVIDIA GPU. The 4-bit weights still need roughly 23 GB. Most developers don't have that machine.
• Needs more guidance. Recursive itself says Flint may need clearer instructions and smaller task breakdowns than the big paid coding assistants. That's the company telling you not to expect Claude Code parity.
• Unverified benchmarks. Every performance claim traces back to Recursive. No independent benchmark lab has confirmed the numbers yet.
• Terminal-Bench gap. Its weakest benchmark — agentic terminal tasks — trails strong dense models by a wide margin.
• The meta-risk. Recursive self-improvement is exactly the capability that worries AI safety researchers, and several bills in Congress have proposed limits on it. Socher says the system stays under human oversight — but this is the first release, not the last.

How Flint compares to other local coding models

The table's message: Flint isn't competing with cloud giants — it's competing for the local-coding lane, where its real rivals are other open-weight models you can run yourself.

Who is behind it, and what it costs them

Recursive is led by CEO Richard Socher, Salesforce's former chief scientist. The company emerged from stealth in May 2026 with $650 million in funding at a $4.65 billion valuation, backed by GV, Nvidia, and AMD Ventures. It also signed a $410 million compute deal with AWS in June — which the company told TechCrunch was "less about headcount and more about agent count." The funding and compute explain how a team small enough to fit in one room could train a 40-billion-parameter model in under four months: money, GPUs, and an AI doing a chunk of the research.

Getting started

The weights are at huggingface.co/RecursiveSI/flint. Recursive is also releasing its own lightweight inference engine, rec, optimized for Flint on M4 Pro and M5 Pro/Max chips, with setup docs for the rest of the stack.

Practical advice straight from the launch post: break complex work into smaller subtasks and give explicit instructions. Treat Flint as a fast, private junior engineer that needs clear specs — not as a drop-in replacement for a frontier cloud model.

FAQ

Is Recursive Flint really free?

Yes — the weights are free to download under the Apache 2.0 license, and there are no API fees because you run it yourself. The hidden cost is hardware: you need a 48GB Apple Silicon Mac or a 48GB NVIDIA GPU.

What hardware do I need to run Flint?

Recursive targets 48GB Apple Silicon Macs (M4 Pro, M5 Pro, M5 Max) or 48GB NVIDIA GPUs. The 4-bit build needs about 23GB of memory. It does not currently target ordinary 8–16GB laptops.

Was Flint actually built by AI?

Not entirely — that's Recursive's marketing frame. The company's own wording is that its Eureka Machine "helped develop nearly every part" of the model-building stack, working alongside a small human team. The automated system proposed architectures, quantization recipes, and kernels, which humans then integrated and shipped. It's the strongest public demonstration of AI-assisted model research so far, but "built by AI" overstates it.

Is Flint better than Claude Code or Codex?

Recursive doesn't claim that — it compares Flint to other open-weight models under 50B parameters, not to frontier cloud assistants. Expect it to be fast and private, but to need more guidance on hard tasks.

Is recursive self-improvement dangerous?

That's the live debate. Some researchers fear AI improving itself could escape human control, and lawmakers have proposed limits. Socher says the Eureka Machine is a supervised research assistant and that any future self-improving version would stay under human oversight with the ability to stop it. Flint itself is a coding model, not a self-improving agent — the recursive loop it demonstrates is still early and human-supervised.

Key takeaways

• Flint is a free, open-weight 40B-parameter coding model that runs locally on a laptop — no cloud, no API fees, no data leaving your machine.
• Its headline claim — a small team plus an AI research system built it from scratch in under four months — is the real story, and the strongest public test of AI-assisted model building yet.
• Company-reported benchmarks are strong for the local-coding lane (74.2% SWE-bench Verified, 125+ tok/s on M4 Pro) but none are independently verified.
• The practical catches: 48GB of hardware, weaker performance on terminal-style agentic tasks, and the need for clearer instructions than paid cloud assistants.
• Recursive's endgame — a research loop that keeps improving itself — is now public and being watched by lawmakers; Flint is explicitly just the starting point.

Published October 22, 2026. Recursive's launch is less than 24 hours old — expect independent benchmarks and more hands-on reports in the coming days, and I'll update this piece as they land.

Sources: Recursive's Flint launch post; Flint model card on Hugging Face; TechCrunch; FourWeekMBA; MarkTechPost; AI Parabellum.

Related on AI Core Updates: Meta's Muse Spark Open: 816B AI Model You Can Download Free; DeepSeek Open-Sources DeepInfer: Its Production Inference Engine, Explained.