“`html
A new deep research agent named MiroThinker-1.7 has been released by the team behind it, featuring a base model built on Qwen3 MoE with 30B total and 3B active parameters for its mini variant.
The release notes highlight benchmark results across various tasks, demonstrating improvements over previous versions such as MiroThinker-1.7-mini (30B/3B active) and other existing models like DeepSeek-V3.2 and GPT-5. Specifically, the 1.7 model scores highest in the BrowseComp-ZH task with a score of 75.3.
- The mini variant is particularly interesting as it only uses 3B active parameters but still performs well compared to other models like Qwen3.5-397B and DeepSeek-V3.2.
- There’s an emphasis on running this model locally, especially the smaller 1.7-mini version, which could provide insights into its performance on consumer hardware.
- The team is also seeking feedback on their context management strategy, including the use of a sliding window with K=5 and episode restarts, as they believe this approach may be more effective for long-context agents.
“`
This brief covers the key points about MiroThinker-1.7’s release and its performance benchmarks, along with an invitation to community feedback on model running conditions.
Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.
