Recursive has released Flint, its first model: an open-weights coding model with 40 billion total parameters but only about 2.5 billion active per token, designed to run coding agents locally. The weights are on Hugging Face under the Apache 2.0 license, and OpenRouter is serving it free at launch. For developers who want an agentic coding model on their own machine without API fees, it is a new small option with strong vendor-reported scores, though none have been independently checked yet.

What Flint is

Recursive cofounder Richard Socher announced Flint on X at 00:29 Beijing time on October 22. According to the Hugging Face model card, Flint is a decoder-only mixture-of-experts model with 39.8 billion total parameters and roughly 2.5 billion active per token, including the output head. Each MoE layer has 256 routed experts, with 6 picked per token plus a shared expert. It takes text in and puts text out, supports reasoning and multi-turn tool calls, and has a stated knowledge cutoff of June 2026.

Recursive says a small team built Flint from scratch in under four months with its automated research system, the Eureka Machine, which helped with data, training recipes, quantization and inference kernels.

Benchmarks, as reported by Recursive

The model card lists these release results, all measured by Recursive:

  • Terminal-Bench 2.1: 54.9%
  • SWE-bench Pro v2: 89.7%
  • SWE-bench Verified: 74.2%

Recursive says it ran Terminal-Bench with Terminus 2 and counted answers found through web retrieval as failures, and ran both SWE-bench tests with mini-swe-agent without internet access. Scores are averages over at least five attempts. Recursive also notes that other developers' scores may not use the same setup, so comparisons with other models should be read with care. The SWE-bench Pro v2 figure is unusually high next to the other two, and no third party has reproduced any of them yet.

Speed and hardware

Recursive 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, using its own optimized inference engine. Cline said in its post that Flint needs only about 24GB of memory. Recursive has not published a memory figure, and the full BF16 checkpoint on Hugging Face is about 80GB. Recursive also publishes quantized versions, which it says score within 0.7 points of the 16-bit model.

On context length, the model card documents 262,144 tokens, which matches OpenRouter's listing. Recursive's blog says Flint supports up to 512K tokens but recommends staying within 256K.

Where to try it

OpenRouter listed Flint at 04:39 Beijing time on October 22 and charges nothing for it at launch. Socher's post also says it can be tried for free in OpenCode and Pi. Free launch access is a common way to get coding models tested, as with DeepSeek V4.1 Flash's free period. Recursive tells users to break complex work into smaller subtasks, since Flint may need more guidance than larger closed models.

Why it matters

Many recent open-weights releases need data-center hardware; Ornith-1.5-122B, a reasoning model that serves on a single 80GB GPU, is one example. Flint aims at laptops. If its scores hold up in independent tests, it would give local setups a credible model for agent work. For how hosted tools compare, see our ranking of the best coding agents.