Supra Reasoning Summarizer — a tiny model to summarize thinking traces from coding agents
这条记录涉及编程工具或代码能力更新，适合开发者评估工作流变化和可复用价值。
SupraLabs 发布了一款名为 SupraLabs/reasoning-summarizer-800m-pre-gguf 的模型，它是一个思维轨迹摘要器。用户或开发者可以向模型发送推理过程和工具调用，模型会生成一个包含标题、副标题、摘要和当前任务的 JSON 格式输出。这使得开发者能够以更具创意的方式向用户展示推理轨迹，而非原始的思维过程。…
It is a thought trace summarizer. Basically, the user/dev can send a reasoning + tool calls (if you want), and then the model will generate a JSON which looks liek this:
https://preview.redd.it/0pqhoqt5khbh1.png?width=1512&format=png&auto=webp&s=0c0a022ec910bc14e6a335ae23581bf2e4c606de
https://preview.redd.it/jouw1fkgkhbh1.png?width=2558&format=png&auto=webp&s=e3087cf2ed2116a4e35790c8f4d17c39d154306c
Devs can implement creative ways to show the reasoning trace to the users with this setup, instead of showing the raw thought trace the model made.
Personally, I like the qwen3.5+ models for this because they create the structure reasoning chains, which you can take to create the thought trace, which looks like how the official qwen chat web app shows the thought trace!
Note: right now, the model is "okay" at summarizing all types of reasoning traces, but we are making an agentic coding specialized model.
Also, I noticed that the model has a somewhat limited vocabulary to start the summary (for the first few words), those will be fixed with the agentic coding thought traces!