Qwen is the most credible open-weight family for teams that want frontier-adjacent quality without an API contract. The Apache-2 licensing on the small and mid-tier sizes is the draw: weights you can self-host, fine-tune, and ship commercially with no royalty strings. The coding and reasoning variants are genuinely competitive, and the size ladder scales the same family from a laptop to a GPU cluster. The catch: the strongest tiers (Qwen3.7-Max, the -Plus multimodal models) are API-only, so the open story tops out below the closed flagship. For self-hosting or multilingual work, solid.
- Open weights: Apache-2 across the small and mid-tier sizes
- Sizes span 0.6B to 235B-plus MoE, from laptop to datacenter
- Strong coding and reasoning variants (Qwen3-Coder, Thinking)
- Leading multilingual coverage across 100-plus languages
- Frequently tops open-model leaderboards for its size class
- Flagship -Max and -Plus tiers are API-only, not open weights
- Largest MoE models need serious GPU memory to self-host
- China-based vendor raises data-residency questions for some teams
- Benchmark leads do not always hold on niche real-world tasks
Qwen is Alibaba’s family of large language models, spanning open-weight releases you can download and self-host and closed flagship tiers served through Alibaba Cloud. It is one of the most widely used open-weight families in production, popular as both a drop-in API alternative and a base for fine-tunes. The draw is permissive Apache-2 licensing on the open sizes paired with quality that tracks close to the closed frontier.
Where it fits
Qwen suits teams that want frontier-adjacent capability without a closed API contract. The open-weight ladder runs from 0.6B dense models that fit on a laptop to large Mixture-of-Experts models for datacenter inference, so the same family covers edge, server, and cluster deployments. Multilingual coverage spans over 100 languages, which makes it a strong default outside English-first stacks. The Qwen3-Coder variants target agentic coding with long context, and the Thinking variants handle math, science, and step-heavy reasoning. It is also one of the most common bases for community fine-tunes, so the surrounding ecosystem of derivatives is deep.
Pricing in practice
The open weights are free. You pay only for the hardware you run them on, whether that is owned GPUs, rented cloud instances, or an inference provider that hosts Qwen behind a per-token API. Apache-2 licensing means no royalties and no usage caps on the open sizes, including commercial use. The hosted flagship tiers (Qwen3.7-Max, the -Plus multimodal models) are priced per token through Alibaba Cloud Model Studio and tend to undercut the closed frontier from OpenAI and Anthropic. The practical cost story is the usual self-host tradeoff: open weights remove per-token fees but add GPU and ops overhead, and the largest MoE models need real memory to serve.
How it compares
Llama, Meta’s open-weight family with a huge tooling ecosystem but a custom license, not Apache-2. Qwen tends to lead on multilingual and coding benchmarks while Llama has broader Western tooling support.
DeepSeek, The other major Chinese open-weight challenger, strongest on reasoning and cost-efficient MoE training. Qwen offers a wider size ladder and more variants; DeepSeek concentrates on fewer, larger reasoning models.
GPT, OpenAI’s closed frontier line. GPT leads at the absolute top end, but it is API-only with no weights. Qwen is the open answer when self-hosting, data residency, or no-contract deployment matters more than the last few benchmark points.
What changed recently
Alibaba shipped the Qwen3.5 series on February 16-17, 2026, led by the open-weight Qwen3.5-397B-A17B MoE model and the hosted Qwen3.5-Plus with a 1M-token context window. The Qwen3.5-Max-Preview placed sixth globally and first in China on LM Arena with an Elo around 1464. The open-weight line continued with Qwen3.6 in April 2026 (the 35B-A3B and 27B models, both Apache-2), and Alibaba introduced the API-only Qwen3.7-Max agent model on May 21, 2026, citing sustained autonomous runs and heavy tool use. By May 2026 the Qwen app reported 234 million users.
Sources
- Qwen3 GitHub repository, github.com, accessed June 2026
- Qwen3.5: Towards Native Multimodal Agents, Alibaba Cloud, February 2026
- Qwen3.6 GitHub repository, github.com, April 2026
- Alibaba introduces Qwen3.7-Max as next-gen AI agent model, TechNode, May 21 2026
- Qwen, Wikipedia, accessed June 2026