Recursive Flint
Quick start (model card): hf download RecursiveSI/flint, build the SGLang image in sglang/, then call the OpenAI-compatible endpoint at http://localhost:30000/v1 with model="flint".
PricePrice not published
Checked22 October 2026
Where it stands
No newer code model in our registry yet.
- Weights
- Weights are public under a permissive licence (Apache-2.0).
How to call it
3 ways to use Recursive Flint.
| Provider | Model id | Endpoint | Docs |
|---|---|---|---|
| Hugging Face | huggingface.co | ||
| Hugging Face (int4 / NVFP4 GGUF) | huggingface.co | ||
| Recursive local engine | docs.recursive.com github.com |
Pricing
Price not published.
No price source linked. Checked 22 October 2026.
Specs
- Input
- Text
- Output
- Text
- Context window
- 262,144 tokens
- Max output
- Not published
- Open weights
- Yes
- Licence
- Apache-2.0
- Training cutoff
- June 2026
- Released
- 21 October 2026
- Status
- Current
- Type
- Code model
What stands out
What its maker marketed at launch, and what people found later, each with its source.
Model largely developed by automated AI research
Recursive says its Eureka Machine helped develop nearly every part of Flint’s model-building stack, from data and training recipes to infrastructure and inference optimization.
Fast local agentic coding
40B-total / ~2.5B-active MoE; 125+ tokens/s on an M4 Pro MacBook Pro and 150+ on RTX Spark; 74.2% SWE-bench Verified, 54.9% Terminal-Bench 2.1 (developer-reported).
Notes
Open-weight coding MoE (39.8B total, ~2.5B active, 262k context). BF16 checkpoint ~80 GB serves on one B300/H200 via the bundled SGLang kit; Transformers modeling code not included at launch. Recursive recommends splitting big tasks into smaller steps. Also free on OpenCode and pi.dev (per Socher).