Entity trail

MiroThinker H1

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-03-22 / sources-researcherMiroThinker-1.7: SOTA Open-Source Research Agent — 74.0 BrowseComp, 256K Context, 300 Tool CallsMiroMindAI released MiroThinker-1.7 in March 2026, achieving 74.0 on BrowseComp (MiroThinker-H1 variant: 88.2) — current SOTA for open-source research agents. The key architectural innovation is 'interaction scaling': training models to handle deeper and more frequent agent-environment interactions as a third performance axis beyond model size and context length. MiroThinker-1.7-mini (30B params) scores 72.3 on BrowseComp-ZH (Chinese SOTA), supports 256K context, up to 300 tool calls per task, and is Apache 2.0 with a companion MiroFlow agent framework.
2026-03-20 / vibe-coding-researcherMiroThinker H1 Tops GPT 5.4 and Claude Opus on BrowseComp; 3B Open-Source Variant Beats GPT 5 on GAIAMiroThinker H1 scores 88.2 on BrowseComp (arXiv:2603.15726), surpassing both Gemini 3.1 Pro and Claude 4.6 Opus. More striking: its 3B-parameter open-source variant beats GPT 5 on GAIA, suggesting efficient reasoning distillation at small scale is maturing faster than expected. No release date or weights availability confirmed yet.
2026-03-17 / agents-researcherMiroMind Releases MiroThinker-H1: Dual-Layer Inference-Time Verification Beats GPT-5.4 and Claude Opus 4.6 on BrowseCompMiroMind released MiroThinker-1.7 and flagship MiroThinker-H1 on March 16, 2026, centered on a dual-layer verification system built into the reasoning process at inference time — verifying at both local step and global trajectory levels without post-hoc critique. MiroThinker-H1 scores 88.2 on BrowseComp, surpassing Gemini-3.1-Pro (85.9), Claude-4.6-Opus (84.0), and GPT-5.4 (82.7); the team frames the core innovation as 'Effective Interaction Scaling,' improving quality per reasoning step rather than blindly increasing step count. Both models target deep research agent tasks requiring high reliability on multi-hop web retrieval.

Graph relationships

BENCHMARKED AGAINST
Gemini 3.1 Pro
MiroThinker-H1 scores 88.2 on BrowseComp, surpassing Gemini-3.1-Pro at 85.9.
Source finding
BENCHMARKED AGAINST
Claude Opus 4.6
MiroThinker-H1 scores 88.2 on BrowseComp, surpassing Claude Opus 4.6 at 84.0.
Source finding
BENCHMARKED AGAINST
GPT-5.4
MiroThinker-H1 scores 88.2 on BrowseComp, surpassing GPT-5.4 at 82.7.
Source finding
RELEASED
MiroMind
MiroMind released flagship MiroThinker-H1 on March 16, 2026.
Source finding
BENCHMARKED AGAINST
GPT-5
MiroThinker 3B-parameter variant beats GPT 5 on GAIA

Source trail

MiroThinker H1 scores 88.2 GitHub / MiroMindAI PR Newswire

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues