Recursive Flint: A New 40B Coding Agent Model Outperforms Qwen
Recursive Flint is a new 40B parameter coding agent model featuring a 2.5B active footprint that outperforms previous models like Qwen3.6-35B-A3B on benchmarks such as TerminalBench 2.1 and SWE-bench Pro v2. It offers a 262k context length and a 3:1 sliding window attention (SWA) to full attention ratio. This model represents a significant step forward for the local LLM community by providing a powerful, efficient tool for coding tasks that is accessible to users with limited hardware resources. It threatens to replace older open models as the go-to option for local coding agents. The model features 40B total parameters with only 2.5B active parameters, a massive 262k token context window, and is supported by Unsloth GGUFs and llama.cpp integrations. It has also received high praise, including coverage by WIRED Magazine.
## BACKGROUND
SWE-bench and TerminalBench are benchmark frameworks used to evaluate the real-world software engineering and terminal-based task capabilities of generative AI models. GGUF is a self-contained file format designed for efficient local inference of large language models on community hardware.