Recursive Releases Flint, an Open-Weight 40B MoE Local Coding Model for Agents with 2.5B Active Params

Recursive Superintelligence has shipped Flint, its first model, and it’s aimed squarely at coding agents that need to stay local. The model is a mixture-of-experts transformer with 39.8B total parameters and about 2.5B active per token. Its weights are on Hugging Face under Apache 2.0.

Flint is text-only, but it is not just a code autocomplete toy. Recursive says it can inspect codebases, fix bugs, implement features and write tests, while also supporting reasoning and OpenAI-compatible tool calls. The company also says its Eureka Machine, an automated research system, helped build almost every part of the stack.

The architecture is unusually specific. Flint has 49 layers in total: two dense layers up front, then 47 MoE layers. Each MoE layer contains 256 routed experts, with a router selecting six per token plus one shared expert that is always active. The model mixes sliding-window attention in 35 layers with full attention in 14 others.

Recursive is also pushing Flint as a practical local model rather than a giant server-only beast. It says 4-bit weights take about 23 GB, and the model can run on 48 GB Apple Silicon Macs — M4 Pro, M5 Pro or M5 Max — or 48 GB NVIDIA GPUs. On the speed side, Recursive claims more than 125 tokens per second on a MacBook Pro and more than 150 on an RTX Spark, with short-context runs that are up to 2x faster than comparable models.

The benchmark picture is mixed, which is probably the honest part. Recursive says Flint leads its chart among MoE models under 50B on agentic coding, with 74.2% on SWE-bench Verified and 95.1% on IFEval. But it trails the strongest dense models, and Terminal-Bench is the sore spot: 54.9 on Flint’s BF16 version, versus 73.0 for dense Qwen 3.8 in Recursive’s own comparison chart.