2.52B dense (2,516,756,480 BF16 parameters; marketed as 2B)
131,072 tokens
Apache 2.0
Open-weight compact code-localization agent built on MiniCPM5-2B and designed to identify the files and line ranges relevant to GitHub issues, bug reports, feature requests, and natural-language questions about software repositories. ColGrep-Agent-2B uses a dense 42-layer Llama-style causal architecture with approximately 2.52B parameters and a 131,072-token context window. Rather than generating patches itself, it explores repositories through three tools: ColGrep semantic code search, a read-only terminal for inspecting candidate files, and a finish tool that submits relevant locations in `path/to/file:START-END` form. The supplied agent harness gives it a ten-turn exploration budget and is intended to let larger coding agents delegate repository search to a small locally hosted specialist. LightOn trained the model with a short supervised cold start using trajectories generated and filtered by Qwen3.5-27B, followed by 500 GRPO reinforcement-learning steps rewarding accurate file- and line-level localization while penalizing unnecessary exploration. Training environments were constructed from SWE-rebench, SWE-Fixer-Train-110K, and additional multilingual repositories paired with issues and bug-fix commits. On LightOn's five-seed code-localization evaluation, the model reports 76.18 File F1 and 19.92 Line F1 on SWE-bench Lite, versus 66.72 and 9.62 for its MiniCPM5-2B base model. On Multi-SWE-bench Flash it reports 62.57 File F1 and 15.93 Line F1, versus 52.01 and 11.31 for the base model. These measurements use 300 instances per benchmark per seed, a ten-turn limit, temperature 0.6, and no reasoning for comparison models; they evaluate localization rather than end-to-end issue resolution. The official release includes approximately 5.04 GB of BF16 safetensors and a separate GGUF repository with multiple quantizations. ColGrep's installer defaults to the approximately 2.5 GB Q8_0 build and runs the agent locally without requiring Python, an external inference server, or an API key.