Flint
Description
Flint is an open language model for coding agents from Recursive, released on 2026-10-21. It is built to investigate codebases, fix bugs, implement features and write tests.
It is a mixture-of-experts model: about 39.8 billion parameters in total, of which roughly 2.5 billion are active for each token, which is why compressed versions can run on a laptop. It reads and writes text, supports reasoning and multi-turn tool calls through an OpenAI-compatible interface, and its documented context is 262,144 tokens. The knowledge cutoff stated by the authors is June 2026.
In the authors' own release evaluations Flint scores 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1; most comparison scores in their table are also reported by developers, so these are not independent checks. The authors note that it needs more guidance than large paid coding assistants, so big tasks are better split into smaller steps.
The weights are public without any acceptance step. The license is Apache 2.0, which allows commercial use; the license covers the weights, while the terms of the developer's hosted service are separate. The repository ships the model, a chat template and a serving kit for SGLang; the full BF16 checkpoint weighs about 80 GB, and quantized variants for NVIDIA and Apple Silicon are published separately. The repository does not yet include Transformers modeling code.
Date of check: 2026-10-22.
Pros and cons
Pros
- Apache 2.0 license, weights open without an acceptance step
- Only about 2.5 billion of 39.8 billion parameters are active per token, so compressed versions suit a laptop
- Reasoning and multi-turn tool calls, documented context of 262,144 tokens
- Strong results on coding tests in the developer's own table
Cons
- The full checkpoint is about 80 GB and needs a server GPU; laptops need the separate quantized versions
- Text only, no image input
- Benchmark figures come mostly from the developer, not from independent tests
- Needs more guidance than large paid coding assistants; big tasks should be split into steps
- No Transformers modeling code in the repository yet
Key facts
- License
- Apache 2.0
- Commercial use
- Allowed
- Hardware requirements
- The full BF16 checkpoint is about 80 GB and is served on one B300 or H200 GPU; the developer prepares separate quantized variants for NVIDIA and Apple Silicon.
- Developer
- Recursive
- Interface languages
- English
- Checked on
- Oct 22, 2026
Editorial score
News about Flint
-
AI startup Recursive launched Flint, its first coding language model developed largely by its internal "Eureka Machine" AI. The 40B-parameter MoE model (2.5B active) runs locally on Apple and Nvidia hardware, hits 125+ tokens/sec on M4 Pro, and is available with open weights on Hugging Face under Apache 2.0. Recursive's long-term goal: AI that builds better AI. #AI #Recursive #Flint #LLM #OpenSource -
Anthropic CEO Dario Amodei urges immediate global speed limits, independent auditors, and shared safety standards on recursive AI self-improvement to prevent systems from outpacing human control and threatening internet stability. #AI safety #AI regulation #recursive self-improvement #Anthropic #Dario Amodei