Recursive releases Flint, an open 40B MoE model for coding agents Other

Recursive Superintelligence released Flint, an open-weights Mixture-of-Experts (MoE) foundation model purpose-built for autonomous coding agents. The model features 39.8 billion total parameters with only 2.5 billion active per token, running under an Apache 2.0 license.

Founded by former Salesforce Chief Scientist Richard Socher, Recursive emerged from stealth in May with $650 million in venture funding. Its technical team assembled researchers from DeepMind, FAIR, Google, and OpenAI, including Vision Transformer author Alexey Dosovitskiy and Tim Rocktäschel. Flint was trained from scratch in less than four months specifically for terminal execution, multi-file code editing, automated debugging, and test generation.

The architecture pairs sliding-window attention with an extreme routing strategy: 256 fine-grained experts per layer, routing each token to six dynamic experts while keeping one shared expert permanently active. Despite its lightweight compute footprint, Flint scored 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1, matching larger 25B to 40B dense models at five times higher throughput.

What it is

• MoE coding agent architecture. Flint contains 39.8 billion parameters total, activating 2.5 billion per token across a 262,144-token context window.
• Expert sparsity. Each transformer layer hosts 256 experts. For each token, the router selects six dynamic experts and routes through one always-active shared expert.
• Benchmark results. Flint achieved 74.2% on SWE-bench Verified and 54.9% on Terminal-Bench 2.1.
• Local execution performance. Generates over 125 tokens per second on Apple M4 Pro MacBooks and over 150 tokens per second on Nvidia RTX Spark GPUs (64GB).
• Inference runtime support. Fully compatible with llama.cpp, Ollama, LM Studio, vLLM, and SGLang, alongside Recursive's custom rec runtime engine for Apple Silicon.
• Permissive licensing. Weights, inference code, and evaluation scripts are released under Apache 2.0.

Model specifications

Sources

• Recursive Official Announcement: Introducing Flint
• Recursive Developer Documentation
• Flint Model Weights on Hugging Face
• Recursive rec Inference Engine on GitHub