Recursive Releases Flint 40B MoE Coding Model

The company states that its automated research system, Eureka Machine, was instrumental in building nearly every part of the model's stack. The model's weights are available on Hugging Face under the Apache 2.0 license.

Flint is a decoder-only MoE Transformer designed for coding agents, capable of investigating codebases, fixing bugs, implementing features, and writing tests. It supports reasoning and OpenAI-compatible tool calls. The model can run on Apple Silicon Macs with 48 GB of RAM or NVIDIA GPUs with 48 GB of VRAM. Using 4-bit weights, it requires approximately 23 GB of storage.

In terms of performance, Flint leads Recursive's internal chart for MoE models under 50 billion parameters on agentic coding tasks, though it trails the strongest dense models. It achieved 95.1% on IFEval and 74.2% on SWE-bench Verified. However, its performance on Terminal-Bench was 54.9%, compared to 73.0% for the dense Qwen 3.8 model.

The architecture of Flint features 49 layers, comprising 2 dense layers followed by 47 MoE layers. Each MoE layer contains 256 routed experts, with a router selecting 6 experts per token, plus one always-active shared expert.

The model utilizes two types of attention mechanisms, with 35 layers employing a 513-token sliding window and 14 layers using full attention without positional encoding. It has 32 query heads and 4 key/value heads. The 2.5 billion active parameter count includes the output head, with the active body alone accounting for approximately 2.2 billion parameters.