A 39.8B-parameter model can be practical on consumer hardware when only about 2.5B parameters are active per token and the weights are quantized. This repository makes Recursive’s coding-agent model available in GGUF formats, shifting the main decision from whether to run it locally to which quality, memory, and context trade-off fits your machine.
Key Capabilities
- The underlying model targets codebase investigation, bug fixing, feature implementation, and test writing, with reasoning and multi-turn tool use. That makes it relevant to agent workflows rather than only code completion.
- Quantized files come in multiple sizes. The documented UD-Q3_K_XL variant is about 18.5 GB and can fit a 24 GB memory setup at shorter contexts; larger variants preserve more of the reference model’s behavior but need more memory.
- The model supports a documented context configuration of 262,144 tokens, but long prompts and reasoning consume memory and output budget. Maximum context capacity does not guarantee uniform performance across long sequences.
- Recursive reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified for its BF16 release. These are reference-model results, not a guarantee that every quantization and runtime will match them.
Who It’s For
Great fit if you want to run a tool-using coding model locally and can choose a quantization to match available memory. Look elsewhere if you need the reference checkpoint’s exact behavior, consistently high accuracy on very long contexts, or a ready-made coding application rather than model weights. Review generated code and constrain tool permissions before allowing consequential actions.
