MiroThinker 1.7 release announcement (MiroMind open research agents)

TL;DR

Tweet from @AdinaYakup (HuggingFace) announcing the MiroThinker-1.7 release on HuggingFace from MiroMind: a 235B “1.7” and a 31B “1.7 mini,” both with a 256K context window and a per-task budget of 300 tool calls, Qwen3-based with a custom agent runtime (MiroThinker-H1). Same lineage as the v1.0/v1.5 paper MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling, pushing the parameter ceiling of the open research-agent family from ~72B/235B-MoE to a denser 235B and dropping the smaller variant to 31B.

Key claims

MiroThinker-1.7 ships in two sizes on HuggingFace: 1.7 (235B params, text-generation) and 1.7-mini (31B). [Collection listing]
Stated per-task budget is 300 tool calls inside a 256K context window. [Tweet body]
Models are Qwen3-based, paired with a custom agent runtime labeled “MiroThinker-H1.” [Tweet body]

Method

The tweet does not contain method details beyond the bullet list above; the HuggingFace collection only lists model artifacts and parameter counts. Background on the post-training recipe (agentic SFT on synthesized multi-hop QA trajectories → DPO with auxiliary SFT loss → online GRPO with streaming rollouts) is documented in the v1.0 paper MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling §4–5.

Results

No benchmark numbers in the tweet or collection. The v1.5 release reported 39.2 / 69.8 / 71.5 / 80.8 on HLE-Text / BrowseComp / BrowseComp-ZH / GAIA-Val-165; v1.7 numbers had not been published at filing time.

Why it’s interesting

This is the third public step in the same MiroThinker scaling story already on the wiki, and the headline number that changes is the per-task tool-call budget: 600 (v1.0/v1.5) → 300 (v1.7). That’s not a typo in the tweet — it’s a regression relative to the paper’s “interaction depth as a third scaling axis” framing in MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling, and it lines up suggestively with the Training-Free Group Relative Policy Optimization finding that better-optimized agents reduce their per-task tool-call count even as accuracy improves (open question already flagged on Tool-Use Agents). Worth watching whether the eventual 1.7 tech report reframes interaction depth as a budget-to-saturate rather than a budget-to-maximize. The 31B-mini variant also lands in the same parameter band as Qwen3-30B-A3B-Thinking-2507 (the v1.5-30B base), so the relevant comparison is whether the recipe transfers to the new base without retuning.