Qwen 3

tl;dr: the Qwen 3 family on Ollama — eight sizes from 0.6B to 235B, every one with tools and thinking modes, and a long-context variant on 4B/30B/235B. The flagship 235B-A22B MoE benchmarks against DeepSeek-R1, o1, o3-mini, Grok-3, and Gemini-2.5-Pro per Alibaba’s own evals.

What it is

qwen3 is the umbrella tag Ollama ships for Alibaba’s Qwen 3 release. The family mixes dense sizes (0.6B, 1.7B, 4B, 8B, 14B, 32B) with MoE sizes (30B-A3B, 235B-A22B). The page lists tools and thinking as capabilities on every variant — meaning all sizes expose function calling and the chain-of-thought reasoning toggle.

The ollama.com page lists 58 tags, 36.7M cumulative pulls, and a last update of roughly 11 months ago for the long-context variants. Default ollama run qwen3 lands on qwen3:8b (5.2GB, 40K context). Two recent additions — qwen3:30b (MoE, 256K) and qwen3:235b (MoE, 256K) — are the headline updates on the README.

Pull command

From the ollama.com detail page:

ollama run qwen3         # defaults to qwen3:8b (5.2GB, 40K context)
ollama run qwen3:30b     # MoE, 19GB, 256K context
ollama run qwen3:235b    # MoE, 142GB, 256K context

The page also lists direct integrations for four agents: ollama launch claude --model qwen3, ollama launch opencode --model qwen3, ollama launch hermes --model qwen3, ollama launch openclaw --model qwen3.

Sizes and hardware

The 30B and 235B are Mixture-of-Experts — activated parameters are 3B and 22B respectively, even though total params are higher. Throughput per token reflects the active count, not the total.

What the README says

Alibaba’s README on the Ollama page describes the Qwen 3 family as the “latest generation of large language models in Qwen series” with reasoning, agent, and multilingual upgrades over Qwen 2.5 / QwQ. Four highlighted axes:

• Reasoning — surpasses QwQ in thinking mode and Qwen2.5 instruct in non-thinking mode on math, code generation, and commonsense logic.
• Human preference alignment — creative writing, role-playing, multi-turn dialogue, instruction following.
• Agent capabilities — tool integration in both thinking and unthinking modes; “leading performance among open-source models in complex agent-based tasks.”
• Multilingual — 100+ languages and dialects, instruction following and translation.

The flagship qwen3:235b is positioned against DeepSeek-R1, o1, o3-mini, Grok-3, and Gemini-2.5-Pro on Alibaba’s own benchmarks. The 30B-A3B is positioned as outcompeting QwQ-32B “with 10 times of activated parameters” — i.e., 3B active vs 32B dense. The 4B dense is positioned as rivaling Qwen2.5-72B-Instruct.

License

The Ollama page does not quote a license in the visible README. Upstream Qwen 3 weights are released under Apache 2.0 — verify at https://github.com/QwenLM/Qwen3 and the model card on Hugging Face before commercial deployment.

Gotchas

• Default qwen3 is the 8B, not the 235B headline. To run the flagship you need to specify :235b and have multi-GPU or cloud hardware ready.
• The 30B and 235B are MoE. The “active” parameter count (3B / 22B) is what determines throughput per token — total parameters determine disk and VRAM.
• Context length varies by size: 40K on 8B / 14B / 32B; 256K on 4B / 30B / 235B. Don’t pick a size by context alone.
• The Ollama page lists tools and thinking as capabilities, but doesn’t detail the tool schema — check the Qwen3 chat template and Hermes function-call format on the upstream repo before wiring tools.
• 235B at 142GB won’t fit on a single consumer GPU. Plan multi-GPU or Ollama’s cloud tier.

Related: DeepSeek-R1 for the open reasoning counterpart, GPT-OSS for the OpenAI-side MoE story, Llama 3.1 for the Meta tier, Ollama for the runtime.

Untested here — facts from the ollama.com /library/qwen3 page fetched 2026-09-10.