Flint is an open weight mixture of experts coding model from Recursive for investigating codebases, fixing bugs, implementing features and writing tests. 39.8B total parameters with about 2.5B active per token · 49 layers with 2 dense and 47 MoE · 256 routed experts, 6 selected plus 1 shared · 262,144 token context · text in, text out · reasoning and tool use · knowledge cutoff June 2026 · Apache 2.0 weights on Hugging Face as RecursiveSI/flint · free OpenRouter endpoint recursive/flint-v1.0:free.
No benchmark data available yet.
- Typetext
- Context262K tokens (~131 books)
- ReleasedOct 2026
- LicenseOpen Source
- StatusActive
Frequently Asked Questions
Related Models
Qwen3 Coder Next FP8 · AlibabaQwen2.5 Coder 32B Instruct AWQ · AlibabaSpeaker Diarization Community 1 · PyannoteJais 3 0.9B · MBZUAIGemma 3 12B (free) · Google DeepMindBenchmarks
Key facts · as of 2026-10-26
- Flint by Recursive. Not enough public benchmark scores to rank yet.
- List price n/a input · n/a output per 1M tokens (as of 2026-10-26).
How to cite · data as of 2026-10-26
Flint · benchmarks, pricing and providers. BenchGecko, data as of 2026-10-26. https://benchgecko.ai/model/flint-v1-0
Credit "Source: BenchGecko" with a link. Prices per provider and Gecko Tests are BenchGecko data (CC BY 4.0); benchmark scores keep their original source, listed in the JSON. JSON · llms.txt · MCP