Recursive Flint: free 40B coding model built to run on a laptop

Key takeaways

• Recursive released Flint, its first model, on 21 October 2026. It is a free coding model with open weights under the Apache 2.0 license on Hugging Face.
• Flint is a mixture-of-experts (MoE) model with about 40 billion total parameters. It uses roughly 2.5 billion of them per token, which is why it can run on a laptop or desktop GPU.
• Recursive reports 74.2% on SWE-bench Verified, 89.7% on SWE-bench Pro v2 and 54.9% on Terminal-Bench 2.1. It also claims more than 125 tokens per second on an M4 Pro MacBook Pro. Recursive ran these tests itself, and no one has reproduced them independently yet.
• When Recursive says Flint was "largely built by AI", it means its automated research system, the Eureka Machine, ran experiments on the model's design, its compression and its kernels. People oversaw that work.
• Recursive says Flint needs more guidance than large closed coding assistants and works best when big tasks are split into smaller steps.

Recursive Flint, released 21 October 2026, is a free coding model with 40 billion parameters under the Apache 2.0 license. Recursive, led by former Salesforce chief scientist Richard Socher, says it runs on a laptop because it uses roughly 2.5 billion parameters per token.

The launch matters for developers who want a coding assistant that works offline, keeps source code on their own machine and has no per-token cloud fees. Most of the headline numbers come from Recursive itself, so this article separates the company's claims from what can be checked today.

What is Recursive Flint?

Flint is a text-only, decoder-only Transformer built for agentic coding: writing and fixing code, running terminal commands and multi-turn tool use. The Hugging Face model card lists 39.8 billion total parameters across 49 layers. Two layers are dense and 47 use a mixture of experts. Each MoE layer has 256 routed experts. A router picks six of them for each token, and a shared expert is always active. In practice only about 2.5 billion parameters do work for any given token. That keeps it fast on consumer hardware while the full model holds far more knowledge.

The model card documents a context window of 262,144 tokens. Recursive's launch post says Flint supports up to 512,000 tokens but recommends staying within 256,000 for the best results. If you are unsure what those limits mean in practice, our explainer on what a context window is covers the trade-offs.

What hardware does Flint need?

Flint will not run in its full-precision form on a typical laptop. The full-precision BF16 checkpoint on Hugging Face is about 80 GB. For servers, the model card recommends one Nvidia B300 or H200 GPU per replica with Recursive's SGLang kit. The laptop claim rests on 4-bit builds. The model card says these are prepared separately for Nvidia and Apple Silicon runtimes, and the Hub links to dozens of community and official quantized versions.

Recursive names three local targets: Apple Silicon Macs with M4 or M5 chips, an M4 Pro MacBook Pro, and Nvidia's RTX Spark desktop with 64 GB of memory. The company does not publish a minimum RAM figure. LM Studio's listing for the GGUF build gives a minimum system memory of 24 GB. According to TechCrunch, Wired's reviewer used a Mac with 48 GB. On that evidence, 16 GB machines look out of reach, and 32 GB or more is the safer assumption for long coding sessions.

Recursive says Flint works with llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang. It also released a small inference engine called "rec", tuned for M4 Pro and M5 Pro/Max chips. If you have not run a model locally before, start with our guide to running a local LLM with Ollama. Check the Hub for the exact quantized build you want before you download it.

How Flint compares on coding benchmarks

Recursive published three headline scores, all measured by the company. The model card says agentic scores are the average pass@1 over at least five attempts. Terminal-Bench was run with the Terminus 2 harness and SWE-bench with mini-swe-agent. Scores for rival models in Recursive's comparison chart were reported by those models' own developers, not re-run by Recursive.

Recursive describes Flint as the best-performing MoE model under 50 billion total parameters on agentic coding benchmarks. It also says Flint rivals dense models with 25 to 40 billion parameters. TechCrunch summarises the comparison this way: Flint beat other models under 50 billion parameters on SWE-bench Verified. On Recursive's average of six tests, only Alibaba's Qwen3.8 27B, a dense model, scored higher. The comparison scores appear only in a chart image, so we do not reproduce them here.

Two caveats apply. First, Recursive counts any Terminal-Bench solution found through web retrieval as a failure, and the model card notes that other developers may not apply the same rule. Second, Recursive says its 4-bit release scores within 0.7 points of the 16-bit model across its benchmarks. That figure is also self-reported. Teams that plan to rely on Flint should test it on their own codebase. Our tutorial on how to evaluate an LLM app in Python shows one way to do that.

Speed claims

Recursive also claims Flint is more than five times faster than dense 25B to 40B models with similar coding scores. That gap is what you would expect from an MoE design: fewer active parameters means less computation and memory traffic for each token generated.

What "largely built by AI" actually means

Recursive presents Flint as an early public test of recursive self-improvement, the idea that AI systems can help build better AI systems. According to the launch post, a small team of people and agents built Flint from scratch in under four months. The company says the Eureka Machine sped up the work by proposing ideas, implementing and testing them, and choosing what to try next. It also explored several directions in parallel and combined the most promising ones.

Recursive lists these specific contributions:

• Model shape: starting from a baseline of 35 billion total and 3 billion active parameters, the system searched expert layouts and arrived at a sparser 40B-total, 2.5B-active design.
• Compression: it found a 4-bit recipe that also quantizes the output layer, which is unusual. It added a distillation stage that, according to the company, nearly halved training steps compared with a standard quantization-aware baseline. TechCrunch reports that low-bit training shrank Flint by 71% with almost no quality loss, according to Socher.
• Kernels: it produced custom low-level code for low-bit maths, expert routing, attention and prompt processing. On Apple's M4 and M5 chips this included a single "megakernel" for the whole decoding step.

What remains unknown: Recursive has not said how much researcher time the system saved, which ideas came from humans and which from the machine, or whether outside researchers can inspect the process. Humans supervised the work. Socher told Wired, as reported by TechCrunch, that the long-term goal is full recursive self-improvement with humans able to stop it. Flint is evidence of AI-assisted model development, not of an AI improving itself unaided.

Company background

Recursive was founded in 2025. According to TechCrunch, it came out of stealth in May 2026 with $650 million in funding at a $4.65 billion valuation. Its backers include GV, Nvidia and AMD Ventures. In June it signed a $410 million computing deal with Amazon Web Services.

What it means

For developers, Flint is one of the strongest claims yet that a useful coding agent can run on a high-end laptop. That matters for privacy-sensitive codebases, offline work and cutting API bills. It fits naturally into local setups such as a local RAG app in Python or a tool-calling agent. Recursive itself says Flint needs more step-by-step direction than large paid assistants, so it complements cloud models rather than replacing them for large refactors.

For the AI industry, Flint is a public test of Recursive's pitch that automated research can let a small team match larger labs. The benchmark figures are self-reported, and the comparison scores came from other developers, so independent leaderboards and hands-on testing will decide how Flint really ranks. The Apache 2.0 licence means those tests can start right away.

Sources

• Recursive: Introducing Recursive Flint (21 October 2026)
• Hugging Face: RecursiveSI/flint model card
• TechCrunch: Recursive launches Flint, a laptop coding model it says was largely built by AI
• LM Studio: Recursive Flint model listing
• Apache Software Foundation: Apache License 2.0