SupraLabs has released a new 800 million parameter model called reasoning-summarizer-800m-pre-gguf, designed to summarize thinking traces from coding agents. The model takes reasoning steps and tool calls as input and outputs a structured JSON object containing a title, subtitle, summary, and current task.
Alongside the model weights, SupraLabs has also released an open-source dataset named reasoning-summaries-61k. The authors note that while the current model is functional for general reasoning traces, they are developing a specialized version for agentic coding tasks to address limitations in vocabulary and structure.