Qwen3

TAS Score: S3/3 — D4/5 — A3 / T2 (D4 not D5: the hosted DashScope API collects usage data and requires an account, and Alibaba is subject to the Chinese regulatory environment. The self-hosted open-weight variant scored here carries none of this.)

⚠️ Two-mode tool. Self-hosted open weights: A3/T2. Hosted DashScope API (Alibaba Cloud Model Studio): A1/T1. This card scores the self-hosted open-weight variant.

Brief Description

Open-weight model family from Alibaba’s Qwen team — dense models from 0.6B to 32B plus Mixture-of-Experts models (30B-A3B and 235B-A22B). Every open-weight size ships under the Apache-2.0 license, so both use and redistribution are uncontested. Switches between a “thinking” mode for reasoning and a “non-thinking” mode for general chat; strong at reasoning, coding, and multilingual tasks (100+ languages).

Small sizes run on consumer hardware; the 235B MoE needs a multi-GPU host. Training data is not published — as with nearly every open-weight model — but the license on the weights is fully OSI-approved. The transparency limit here is the training corpus, not usage rights.

Architectural Role

Compute/inference layer: general-purpose open-weight LLM run on your own hardware. Local alternative to cloud chat and reasoning APIs across the full size range — a 0.6B model on a laptop up to a 235B MoE on a GPU server.

Technical Autonomy

  • Works without internet (after model download)
  • Stores data locally
  • Does not require external accounts (self-hosted)
  • Allows data export — Apache-2.0, standard HuggingFace format
  • Hosted DashScope API requires an account and sends data to Alibaba servers

Philosophical Assessment (whose.world criteria)

Criterion Status Comments
Pause ✅ Stop inference. Model weights stay on disk.
Exit ✅ Standard model format. Apache-2.0 — no restrictions on use or redistribution.
Recoverability ✅ Re-download from HuggingFace / ModelScope or restore from backup.
Visibility ✅ Apache-2.0 across all open sizes. Open weights. Training data not published.
External Dependencies ⚠️ Self-hosted: none. Hosted API: Alibaba infrastructure (China-based).

Configuration (Minimal) 

# Consumer hardware — dense model via Ollama
ollama run qwen3

# Server — 235B MoE via vLLM
vllm serve Qwen/Qwen3-235B-A22B --tensor-parallel-size 8

# Or via SGLang
python -m sglang.launch_server --model Qwen/Qwen3-32B

The full range is Apache-2.0, so the same license applies whether you run the 0.6B model or the 235B MoE — only the hardware requirement changes.

Alternatives

Alternative Autonomy Notes
DeepSeek-R1-Distill-Qwen A3 / T2 Reasoning distill built on a Qwen base. Apache.
MiroThinker A3 / T2 Verification-centric reasoning. Apache-2.0.
Qwen-Max A0 / T0 API-only flagship. Proprietary, no weights.

Trajectory

Direction: mixed

Alibaba runs one of the most aggressive open-weight strategies of any major lab — every core Qwen3 size, including the 235B flagship MoE, is Apache-2.0. But Qwen is a corporate project from a Chinese company subject to its regulatory environment, and the top-tier Qwen-Max line is kept API-only and closed.

Signal assessment:

Signal Status Evidence
License ✅ Apache-2.0 across all open-weight sizes (0.6B–235B).
Feature gating ⚠️ Flagship Qwen-Max is API-only — not released as weights.
Self-hosting ✅ Standard formats. Ollama, vLLM, SGLang, llama.cpp compatible.
Governance ⚠️ Corporate (Alibaba). Chinese regulatory environment.

Signal key: ✅ opening · ➖ neutral · ⚠️ closing

Sources

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