Most agent frameworks chase better benchmarks by scaling the model. MiroThinker's wager is that the cheaper, less-explored lever is interaction depth — how many times an agent can search, read, and run code before it loses the thread. It treats agent-environment interactions as a third performance dimension alongside parameter count and context length, and trains specifically to stay coherent across long tool-use chains.
What Sets It Apart
- Interaction as a scaling axis. Sustains up to 300 tool calls per task within a 256K-token context, so research problems that need dozens of lookups don't get truncated halfway through.
- Open weights at two scales. Ships as 30B (mini) and 235B variants on Hugging Face under Apache 2.0, letting you trade cost for capability instead of being locked to one size.
- Competitive where it counts. Reports 82.7% on GAIA-Val-165, 74.0% on BrowseComp, and 75.3% on BrowseComp-ZH — close to commercial deep-research systems while staying fully inspectable.
- Tool stack built in. Web search, content extraction, and code execution are integrated rather than bolted on, with context management that favors recent observations.
Who It's For
Great fit if you're building agentic research, web-navigation, or prediction systems and want open weights you can self-host, fine-tune, and audit — especially for mixed Chinese-and-English browsing tasks. Look elsewhere if you need a lightweight, low-latency assistant: a model that expects hundreds of tool calls and a 256K context is heavy for simple chat or single-shot Q&A, and the 235B variant demands a serious GPU budget.
