Release of RecursiveSI Flint Model
RecursiveSI has announced the release of Flint, a new large language model specifically designed for coding agent tasks. The model features 40 billion parameters, utilizing a Mixture-of-Experts (MoE) architecture with 2.5 billion active parameters. According to technical specifications hosted on Hugging Face, the model supports a context window of 262,000 tokens and employs a 3:1 Switched Weight Attention (SWA) to full attention ratio.
Performance Benchmarks and Technical Specifications
In preliminary performance data released by the developer, Flint demonstrates significant capabilities in software engineering benchmarks. The model reportedly outperforms the Qwen3.6-35B-A3B model in TerminalBench 2.1 and SWE-bench Pro v2. While some users on community forums like Reddit have questioned the choice of TerminalBench 2.1 over more recent versions like 4.0, the reported data suggests Flint may offer a competitive alternative for users with limited GPU resources.
The model is available in various formats, including GGUF versions provided by Unsloth, and has received pull requests for integration into the llama.cpp library. Technical discussions on Reddit have noted the model’s performance using NVFP4 quantization, with some users observing that it performs closely to the BF16 version.
Community Reception and Early Testing
Initial feedback from early adopters has been mixed, ranging from high praise to skepticism regarding the source of the performance claims. Some users have reported that the model exhibits advanced agentic behavior, such as attempting to read non-existent files during testing—a behavior some users noted was absent in other models like Ornith 1.5.
However, some community members have expressed caution, questioning whether the high levels of enthusiasm on social media platforms are organic or “astroturfing.” There is also ongoing debate regarding the practical deployment of the model on consumer-grade hardware, such as NVIDIA 3090 and 4090 GPUs, due to the specific memory requirements of its architecture.
While some early testers have described the model as “scary good,” citing coverage in WIRED Magazine, these assessments remain unverified by independent third-party audits. As with many new releases in the open-source AI space, consistent quality may vary depending on the specific quantization method and KV cache settings used by the end user.