Flint Coding Model: Recursive’s Fast 40B Open Model Explained
The Flint coding model is a new open coding model from Recursive, and it is built to be fast enough to run on a laptop. It has 39.8B total parameters but only about 2.5B are active per token, which is why it is so quick. Recursive says it leads mixture-of-experts models under 50B on agentic coding. In the video below I test it on real coding work. Here is what it is, who made it, how it scores and where you can get it.
Key takeaways
Flint is made by Recursive and was built in under four months with help from its own AI research system.
39.8B total, about 2.5B active, Apache 2.0 licence.
Over 125 tokens a second on an M4 Pro MacBook Pro, per Recursive.
Who made Flint?
Flint comes from Recursive Superintelligence, Inc. Its launch blog, dated 23 October 2026, says it was built by “a small team of people and agents” in under four months.
The agents part is the interesting bit. Recursive used what it calls the Eureka Machine, a system for large-scale automated AI research and development. It suggests ideas, builds them and tests them. Recursive says it helped with the training data, optimizers, training recipes, infrastructure and inference speed-ups. It started from a 35B/3B baseline and searched for a better design.
35 sliding-window layers (window 513) + 14 full attention
Experts
256 experts, top-6 per token, plus a shared expert
Vocabulary
150,144
Knowledge cutoff
June 2026
Licence
Apache 2.0
Recursive also says a few training tricks made a big difference. Offline distillation from a 16-bit teacher nearly halved the training steps needed compared with its baseline. Its 4-bit NVFP4 version scores within 0.7 points of the 16-bit model. It also wrote custom kernels, including a Metal megakernel for Apple M4 and M5 chips.
Want fast open coding models working in your business? Inside the AI Profit Boardroom you get the Agent OS — Hermes, OpenClaw, OpenCode, Claude and Codex in one place — plus step-by-step Hermes training and four live coaching calls a week.
This one confuses people, so here is the honest answer. The model card lists a context length of 262,144 tokens. The launch blog says Flint supports up to 512K, but recommends staying within 256K. The Ollama listing shows 256K too.
In practice, plan for about 256K tokens. That is what the model card and Ollama list, and it is what Recursive recommends. Treat 512K as a stretch limit, not a daily setting.
These are Recursive’s own numbers from the model card. Terminal-Bench was run with Terminus 2, SWE-bench with mini-swe-agent, and each score is an average pass@1 over at least five attempts.
Benchmark
Flint
SWE-bench Verified
74.2%
SWE-bench Pro v2
89.7%
Terminal-Bench 2.1
54.9%
Recursive says Flint leads MoE models under 50B on agentic coding and rivals dense 25–40B models while being more than 5x faster. These are vendor numbers, so test it on your own code before you swap anything important over.
How fast is it?
Speed is the main selling point. Recursive’s blog says Flint runs at over 125 tokens a second on an M4 Pro MacBook Pro and over 150 tokens a second on a 64 GB NVIDIA RTX Spark. It says that is up to 2x faster than comparable models on Apple silicon and RTX Spark.
On data-centre hardware, the vLLM blog reports about 500 tokens a second for a single request on a GB200 at 100K context. Recursive’s own Mac runtime, rec, claims 1.4–2.6x the throughput of llama.cpp on an M4 Pro with 48 GB.
Recursive recommends temperature 1.0, top-p 0.95 and top-k 20. Reasoning is always on. Give it at least 32,768 output tokens for hard problems; the OpenCode guide uses a 65,536 budget. It supports OpenAI-style tool calls, so it plugs into most coding agents.
Why pick Flint over a bigger open model? It comes down to speed and hardware. Most strong open coding models today are large: GLM-5.3-Flash has 320B total parameters and Step 5 Preview has 600B. You can’t run those on a laptop. Flint has 39.8B total and about 2.5B active, so a 4-bit copy is about 23 GB and fits on a 48 GB Mac or GPU.
The trade-off is that the biggest models still win on the hardest tasks. A good way to use Flint is as your fast, private, everyday coder for edits, tests, refactors and quick questions, and to hand the really hard problems to a bigger cloud model when you need to.
Choose Flint when you want speed, privacy and no token bill.
Choose a big cloud model for very large codebases or very hard reasoning.
Use both in a coding agent that lets each agent pick its own model.
Who should use Flint?
Flint suits solo developers, small teams that can’t send code to a third party, and anyone who already owns a 48 GB Mac or GPU. If you don’t have that hardware, you can still try it through OpenCode’s free tier or OpenRouter before you decide whether local is worth it.
My verdict on the Flint coding model
Flint is a strong open model that is small enough and fast enough to run on a good laptop, and it carries a simple Apache 2.0 licence. If you want a coding model you control, on your own machine, the Flint coding model is one of the best options right now. Just plan for 256K context and check the benchmarks on your own work.
About Julian Goldie
I’m Julian Goldie — founder of Goldie Agency and the AI Profit Boardroom. I test new AI agents almost every day for 432K+ YouTube subscribers and run Hermes, OpenClaw, OpenCode, Claude and Codex side by side in my own Agent OS. Claim your free strategy session or join the AI Profit Boardroom.
FAQ: flint coding model
Who made the Flint coding model?
Recursive Superintelligence, Inc. It was built in under four months by a small team working with Recursive’s automated research system, the Eureka Machine.
How big is Flint?
39.8B total parameters with about 2.5B active per token, using 256 experts with top-6 routing plus a shared expert.
What is Flint’s context window?
The model card lists 262,144 tokens. Recursive’s blog says up to 512K but recommends staying within 256K.
Is Flint open source?
The weights are released under the Apache 2.0 licence on Hugging Face.
Can I try Flint for free?
Yes. OpenCode’s free tier offers Flint-40B with rate limits, OpenRouter lists a free version, and you can run it locally at no cost.
Ready to put Hermes to work in your business? Join the AI Profit Boardroom for the Agent OS, setup training and daily updates.
