MiroThinker: MiroMind's Qwen3-Based Open-Source Deep Research Agent

MiroThinker is an open-source deep research agent from MiroMind AI, a Singapore-based lab founded by ex-Alibaba DAMO Academy researchers (co-founder Song Bai, with Lidong Bing previously leading DAMO research). It is not a general-purpose LLM. It is a fine-tuned Qwen model specialized for web-search-heavy, tool-calling research tasks, released under Apache 2.0. ## What it actually is MiroMind ships four components: MiroThinker (the agent model weights, fine-tuned on Qwen3 and Qwen2.5 bases), MiroFlow (the orchestration runtime), MiroVerse (a 147K-trajectory training dataset), and dr.miromind.ai (a hosted product). Recent releases include v1.5 (Jan 2026, 30B + 235B on Qwen3-30B-A3B and Qwen3-235B-A22B-Thinking-2507), v1.7 (Mar 2026, 30B mini + 235B, 256K context, 300 max tool calls), and a proprietary H1 variant. Notably, the advertised tool-call ceiling has dropped across versions (600 in v1.0 → 400 in v1.5 → 300 in v1.7), suggesting early numbers were aspirational. ## Interactive scaling claim MiroMind frames agent improvement along three axes: parameter count, context length, and agent-environment interaction depth. They argue MiroVerse's multi-step trajectory dataset makes them the first open stack to systematically train on the third axis. Frontier closed agents (OpenAI Deep Research, Anthropic and Gemini Deep Research) clearly use similar trajectory data internally, so the open-source release is the real contribution; the 'third scaling axis' framing is marketing more than paradigm shift. ## Benchmarks, honestly Headline number: 88.2 on BrowseComp (H1) versus Claude 4.6 Opus 84.0, Gemini 3.1 Pro 85.9, and GPT-5.4 82.7. v1.7 hits 74.0. BrowseComp tests many-step web search, evidence gathering, and citation discipline — exactly what MiroThinker was trained for. v1.7 also sets a new open-source SOTA on BrowseComp-ZH (75.3) and posts 82.7 on GAIA-Val-165. The model is not benchmarked on general coding, math, or creative writing, and hands-on reviewers report the 30B mini struggles outside its research niche. Rule of thumb: if the task is multi-step web research with citations, MiroThinker belongs in the conversation; otherwise it does not. ## Deployment reality MiroThinker is an orchestrator, not a self-contained model. The README requires E2B (sandboxed code execution), Serper (Google Search API), Jina (reranker, embeddings, URL reader), and a summarization LLM endpoint. Inference runs on local GPU — Qwen3-30B-A3B needs ~40GB VRAM, the 235B variant needs roughly 8×H100 — but every search, scrape, and code call hits paid third-party APIs, and a typical deep-research task is 50-100 tool calls. MiroMind's marketed '$0.07/call vs OpenAI Deep Research $2/call' covers inference only and elides external API spend. ## Verdict The Apache 2.0 license is genuinely clean, the Qwen3-base strategy is sound, and the BrowseComp results are real. What is overstated: 'beats Claude/Gemini/GPT-5.4' is benchmark-specific, the 600-tool-call narrative no longer applies, and cost-per-call comparisons hide operational overhead. Community traction is quiet — two HN threads at 1 point with 0 comments, ~8K GitHub stars. Target user is teams building research-agent pipelines on open weights, not casual chat or coding users. Easiest entry path is the hosted dr.miromind.ai before committing to self-hosting. Sources: GitHub MiroMindAI/MiroThinker (Apache 2.0, ~8,128 stars), Hugging Face miromind-ai org, arXiv 2511.11793 (v1.0 tech report), arXiv 2603.15726 (1.7 and H1 tech report), arXiv 2602.22808 (MiroFlow paper), and the MiroVerse-v0.1 dataset.