MiroThinker

TAS Score: S3/3 — D4/5 — A3 / T2 (D4 not D5: hosted app (dr.miromind.ai) builds user profiles and requires account — trajectory mixed due to commercial push alongside open weights.)

⚠️ Two-mode tool. Self-hosted: A3/T2. Hosted app (dr.miromind.ai): A1/T1. This card scores the self-hosted open-weight variant.

Brief Description

Open-weight reasoning model from MiroMind (30B and 235B parameter variants). Designed for deep research and multi-step verification — proposes hypotheses, searches for evidence, identifies mismatches, revises conclusions. Runs locally via vLLM or SGLang. Hosted version available but requires account.

Architectural Role

Compute/inference layer: reasoning model for complex research and analysis tasks. Replaces hosted research agents (OpenAI Deep Research, Perplexity Pro) when run locally. The 30B variant runs on a single high-end GPU; the 235B MoE variant (22B active) requires multi-GPU setup.

Technical Autonomy

  • Works without internet (after model download)
  • Stores data locally (model weights are standard files)
  • Does not require external accounts (self-hosted variant)
  • Allows data export — standard HuggingFace model format
  • Hosted app requires account and sends data to MiroMind servers

Philosophical Assessment (whose.world criteria)

Criterion Status Comments
Pause ✅ Stop inference server. Model weights stay on disk.
Exit ✅ Model weights are downloadable files. Use with any vLLM/SGLang compatible setup.
Recoverability ✅ Re-download weights from HuggingFace or restore from backup.
Visibility ✅ Apache-2.0. Open weights on HuggingFace. Training code on GitHub.
External Dependencies ⚠️ Self-hosted: none. Hosted app: depends on MiroMind infrastructure.

Configuration (Minimal) 

# Using vLLM (requires NVIDIA GPU with sufficient VRAM)
pip install vllm
vllm serve miromind-ai/MiroThinker-v1.5-30B \
  --tensor-parallel-size 1 \
  --max-model-len 65536

# Using SGLang
pip install sglang
python -m sglang.launch_server \
  --model-path miromind-ai/MiroThinker-v1.5-30B \
  --tp 1 --host 0.0.0.0 --port 1234

# Via Ollama (if GGUF quantization available)
ollama pull mirothinker:30b
ollama run mirothinker:30b

Alternatives

Alternative Autonomy Notes
DeepSeek-R1 A3 / T2 Open-weight reasoning. Distilled versions run via Ollama.
Ollama + Qwen3 A3 / T2 Simpler setup. Less specialized reasoning.
OpenAI Deep Research A0 / T0 Cloud-only. Proprietary. All data on OpenAI servers.

Trajectory

Direction: mixed

Open weights published on HuggingFace (Apache-2.0). Active research team. But commercial direction is visible: hosted app with account requirement, enterprise pitch for financial services. Classic pattern of open-core with closing commercial layer.

Signal assessment:

Signal Status Evidence
License ✅ Apache-2.0 for model weights. Open.
Feature gating ⚠️ Hosted app may get features not available in open weights. Enterprise tier emerging.
Self-hosting ✅ Standard HuggingFace distribution. vLLM/SGLang compatible.
Governance ⚠️ Corporate-led (MiroMind Inc). Small team, venture-backed.

Signal key: ✅ opening · ➖ neutral · ⚠️ closing

Sources

This site uses Just the Docs, a documentation theme for Jekyll.