Recursive releases Flint, a free open-weights coding model built to run on a laptop
Announced · Published · By Contractor AI Shop
Recursive released Flint, a mixture-of-experts coding model with 40B total and about 2.5B active parameters, free under Apache 2.0 on Hugging Face. It ships MLX INT4 weights for Apple Silicon and an NVFP4 GGUF for NVIDIA hardware, so it runs offline on your own machine.
Key facts
- Recursive claims 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. Those speeds are Recursive's. Recursive
- Recursive says Flint works with llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang, and it released rec, its own inference engine tuned for M4 Pro and M5 Pro/Max Macs. Recursive
- The MLX INT4 weights take about 22.6 GB, per the Hugging Face model card. Recursive (Hugging Face)
- The NVFP4 GGUF file is 22.9 GB; Recursive recommends at least 48 GB of GPU memory, or unified memory on Spark, and says it was tested on DGX Spark and RTX Spark. Recursive (Hugging Face)
- WIRED's Will Knight tested Flint on a Mac with 48 GB of RAM and wrote that it nailed his coding requests every time. WIRED (Will Knight)
- TechCrunch reports Recursive built Flint in under four months with heavy help from its automated research system, the Eureka Machine, and that the company raised $650 million at a $4.65 billion valuation. TechCrunch (Aditya Mehta)
Our take: what it means for contractors
A strong new candidate for our local-AI pages: the Mac build is about 22.6 GB, and WIRED ran it on a 48GB Mac. It's a coding model, so it's for building tools and scripts, not writing bids. Recursive itself says it may need more guidance than the big paid coding assistants.
What we don't know yet
- Independent benchmarks (all scores and speeds are Recursive's)
- Speeds at long contexts
- How it compares with other local models, such as DeepSeek V4 Flash, on the same Mac
- oMLX support (Recursive names llama.cpp, Ollama, LM Studio and its own rec engine)
Sources
- Recursive · Introducing Flint · 2026-10-21 · Primary
- Recursive (Hugging Face) · RecursiveSI/flint-int4 · Primary
- Recursive (Hugging Face) · RecursiveSI/flint-nvfp4-gguf · Primary
- WIRED (Will Knight) · This Startup Is All In on Recursive Self-Improvement. Its Latest Model Is Scary Good · 2026-10-21 · Secondary
- TechCrunch (Aditya Mehta) · Recursive launches Flint, a laptop coding model it says was largely built by AI · 2026-10-21 · Secondary