Recursive Flint: Can a Free Coding Model Run on Your Laptop?

Recursive Flint arrived on October 21, 2026, and its pitch is refreshingly simple: a capable coding model you can download for free and run on your own computer. Flint is the first public model from Recursive, the AI research startup led by Richard Socher, and it was released under the permissive Apache 2.0 license on Hugging Face. For developers who want an agentic coding assistant without sending source code to a cloud API, this is one of the most interesting open-weight releases of the autumn.

• Size: 39.8 billion total parameters, with about 2.5 billion active per token, in a mixture-of-experts design
• Context: 262,144 tokens, with a stated knowledge cutoff of June 2026
• Scores (Recursive's own): 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified
• Speed (Recursive's claim): over 125 tokens per second on a MacBook Pro with an M4 Pro chip

What Is Recursive Flint?

Flint is a decoder-only mixture-of-experts transformer built specifically for coding agents. According to its model card, it is designed to investigate codebases, fix bugs, implement features and write tests, and it supports reasoning plus multi-turn tool use through OpenAI-compatible tool calls.

The architecture is worth a closer look. Each of its 47 MoE layers holds 256 routed experts, and only six are selected per token alongside an always-active shared expert. That is why a model with nearly 40 billion parameters behaves, in compute terms, more like a 2.5 billion parameter one. Recursive also mixes 35 sliding-window attention layers with 14 full-attention layers, a common trick for keeping long-context inference affordable.

Built With an Automated Research System

Recursive says a small team trained Flint from scratch in under four months, leaning heavily on its automated research platform, which the company calls the Eureka Machine. The model went through pretraining, mid-training, supervised fine-tuning and reinforcement learning. Crypto Briefing reports that Recursive has raised $650 million at a $4.65 billion valuation, with backers including GV, Greycroft, NVIDIA and AMD Ventures.

Can You Run Flint Locally?

Yes, with the right hardware. The full BF16 checkpoint is about 80 GB, and Recursive's official SGLang serving kit targets a single B300 or H200 GPU per replica. For everyday machines, the model card says quantized variants are prepared separately for NVIDIA and Apple Silicon runtimes, and the Hugging Face page already links quantized builds for llama.cpp, Ollama and LM Studio.

Crypto Briefing reports that Recursive describes Flint as workable for agentic coding on consumer machines with 48GB of RAM. That puts a local coding agent within reach of a well-equipped laptop or a compact AI workstation, which pairs nicely with the high-memory mini PCs we cover in our AI hardware and model coverage.

How Good Is Flint at Coding?

Recursive's numbers are strong for a model this size, but they are the company's own. The model card notes that its evaluations use pass@1 averaged over at least five attempts, and that Flint's Terminal-Bench score treats solutions found via web retrieval as failures, a stricter rule than some competitors apply. RuntimeWire also points out that the speed and benchmark figures have not been independently verified yet, so treat them as a promising starting point rather than a final verdict.

Recursive is candid about limits too. The model card recommends testing generated code, reviewing consequential changes, and having tool-using applications enforce their own permissions. That is sensible advice for any coding agent, local or hosted.

Why Local Coding Models Matter

The appeal of a model like Flint is privacy and predictability. Code never leaves your machine, there is no per-token bill for a long refactoring session, and teams with strict data policies get a capable agent they can audit. Flint joins a growing lineup of efficient open models, including Liquid AI's compact agent model covered in our look at LFM2.5 for local agents. Developers can download Flint from Hugging Face today and put its claims to the test on their own repositories.

Sources: Recursive Flint model card on Hugging Face — October 21, 2026; Crypto Briefing — October 21, 2026; RuntimeWire — October 21, 2026.