REDDIT · REDDIT// 31d agoMODEL RELEASE

MiroThinker 1.7 lands for deep research

MiroMind has released the open-source MiroThinker-1.7 family in 30B and 235B sizes, positioning it as a deep-research agent model for long-horizon tasks rather than general chat. The new models ship with 256K context, support up to 300 tool calls, and are backed by a broader agent stack that emphasizes stepwise verification and benchmarked research workflows.

// ANALYSIS

This looks like a serious attempt to turn “agentic reasoning” from a demo trope into an engineering discipline with benchmarks, tooling, and deployment guidance.

–The big story is not just model size; it is the packaging of model, tool-use workflow, and evaluation configs into a reproducible research-agent stack
–A 256K window plus hundreds of tool calls targets BrowseComp-style deep-search tasks that most open models still struggle to handle reliably
–The 30B mini variant matters because it suggests MiroMind wants developer adoption, not just leaderboard bragging rights on giant checkpoints
–The verification-centric architecture pitch is notable because it directly addresses the usual agent failure modes: fake fixes, broken chains of thought, and weak rollback discipline

// TAGS

mirothinkerllmagentreasoningopen-sourcebenchmarksearch

DISCOVERED

31d ago

2026-03-11

PUBLISHED

31d ago

2026-03-11

RELEVANCE

8/ 10

AUTHOR

External_Mood4719