flint-GGUF: 40B MoE Coding Agent Achieves 89.7% on SWE-bench

Flint, developed by Recursive, is a language model specialized for coding agents. It is designed to handle the entire development workflow, including codebase analysis, bug fixing, feature implementation, and test writing. It adopts a mixture-of-experts structure where only about 2.5B of its total 39.8B parameters are activated per token.

1 / 4

This repository provides the GGUF quantized version offered by unsloth. It includes quantization files of various bit widths, from the BF16 original to Q2_K_XL, Q4_K_XL, and Q8_0, allowing selection based on hardware specifications. It is compatible with major inference engines such as llama.cpp, Ollama, and vLLM, and supports tool use through an OpenAI-compatible API.

Performance metrics recorded 89.7% on SWE-bench Pro v2, 74.2% on SWE-bench Verified, and 54.9% on Terminal-Bench 2.1. It supports a long context window of 262,144 tokens and is designed to maintain reasoning processes and tool call results in the conversation history, making it suitable for complex agent tasks.

Distributed under the Apache-2.0 license, it allows for commercial use. It is useful for building coding assistant agents in local environments or for reducing cloud dependency. Since memory requirements vary depending on the quantization level, you should select the file version that matches your available GPU or CPU resources.