Recursive
Recursive opens the weights of Flint, a coding model for local use
Recursive released Flint, a mixture-of-experts coding model with open weights under the Apache 2.0 license, and Ollama already carries it.
Recursive released the weights of Flint, a coding model, under the Apache 2.0 license, according to its Hugging Face model card. Developers can download it and run it on their own hardware.
The model card was created Oct. 21, 2026. Ollama already carries Flint.
Flint Picks 6 of 256 Experts for Each Token
Flint picks 6 of 256 routed experts in each layer for every token. Only about 2.5 billion of its 39.8 billion parameters are active at once. A shared expert always runs. Flint is a mixture-of-experts model that takes text in and gives text out. It supports reasoning and multi-turn tool use through OpenAI-compatible tool calls.
The model card documents a context window of 262,144 tokens. Recursive’s post says Flint supports up to 512,000 tokens and recommends staying within 256,000 for the best results.
Recursive Reports 74.2% on SWE-bench Verified
Recursive reported 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified in its own release evaluations on Oct. 21, 2026. Each is an average pass@1 over at least five attempts. Terminal-Bench ran with Terminus 2 and SWE-bench with mini-swe-agent.
Recursive counts Terminal-Bench solutions found through web retrieval as failures and says developer-reported scores for other models may not apply the same adjustment. The model card says evaluation setups vary across sources.
Recursive also claims Flint leads mixture-of-experts models under 50 billion total parameters on agentic coding benchmarks. It says Flint rivals dense models of 25 billion to 40 billion parameters while running more than five times faster. On Recursive’s own inference engine, it says, Flint generates more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on a 64GB NVIDIA RTX Spark at short contexts, up to twice as fast as comparable models.
Ollama Added Flint Support on Release Day
Ollama lists Flint as flint at 40 billion parameters, 23GB and 256K context. Its v0.40.8 release, dated Oct. 21, 2026, added Flint support on its MLX backend. The Ollama page shows launch commands for Claude Code, OpenCode, Hermes Agent and OpenClaw with --model flint.
The BF16 checkpoint on Hugging Face is about 80GB. Recursive’s SGLang serving kit runs on one B300 or H200 GPU per replica. Quantized variants for NVIDIA and Apple Silicon are prepared separately. Recursive says the NVFP4 version scores within 0.7 points of the 16-bit model across its benchmarks.
Recursive Pitches Local Use and Credits Automated Research
Recursive pitches local use as a way to work offline, keep data on the computer and avoid API fees.
Recursive says a small team of people and agents built Flint from scratch in under four months. Its automated research system, the Eureka Machine, contributed to data, training recipes, infrastructure, quantization and inference work.
Recursive Advises Splitting Complex Work
Recursive tells developers to break complex work into smaller subtasks with explicit instructions. Flint may need more guidance than larger closed-source coding models, it says. The model card says applications that let the model use tools should enforce their own permissions and execution boundaries.
Our View
We think the caveat about guidance decides who gains from this release. Developers who point Claude Code or OpenCode at Flint through Ollama stop paying API fees and keep code on the machine. They also become the ones who split large tasks into subtasks and write explicit instructions.
The saving lasts only while that extra work costs less time than the fees avoided.