R
Flint
Flint is an open-weight mixture-of-experts model from Recursive, with about 2.5B active parameters out of 40B total, built by Recursive's autoresearch system Eureka Machine for coding agents.
Specifications
| Lab | Recursive |
|---|---|
| Context window | 262.1K tokens |
| Input price | $0.00/1M |
| Output price | $0.00/1M |
| Release | Oct 2026 |
Strengths
- Leads MoE models under 50B total params on agentic coding (74.2% SWE-bench Verified, 54.9% Terminal-Bench 2.1)
- Only ~2.5B active params per token: 125+ tok/s on an M4 Pro MacBook, offline on local hardware
- Apache 2.0 open weights with SGLang, vLLM, Ollama, and llama.cpp support
- 262K context window for large codebase work
Weaknesses
- Coding-specialized: weaker on general reasoning and non-technical tasks
- First model from a new lab with limited real-world validation
- ~80GB BF16 checkpoint needs serious VRAM/RAM despite low active params
Best for
- Local agentic coding: codebase investigation, bug fixes, feature work, tests
- Offline or privacy-sensitive development without API fees
- Self-hosted code assistants on Apple Silicon or single-GPU rigs
In Depth: Flint
Summary
Flint is an AI model from Recursive.
Released in Oct 2026. It takes text input and produces text output, with a context window of 262.1K tokens. Input pricing is $0.00 per 1M tokens and output is $0.00 per 1M tokens on OpenRouter.