Qwen

Qwen is not one model but an umbrella maintained by the Qwen Team at Alibaba Cloud, spanning the GitHub org QwenLM: general-purpose LLMs (Qwen3, currently at generation Qwen3.8), a dedicated coding CLI (qwen-code, a former Gemini CLI fork now independently developed), vision-language models (Qwen3-VL), image generation/editing (Qwen-Image), omni and audio models (Qwen3-Omni, Qwen3-ASR, Qwen3-TTS), embeddings, an agent framework (Qwen-Agent), and a free consumer chat app (chat.qwen.ai). Most of these are separate, independently active repos rather than one monolith.

The mainline Qwen3 model sizes are released under Apache 2.0 (confirmed in the QwenLM/Qwen3 README's license section). But as of August 2026, Alibaba started releasing its Max-tier flagship weights too (Qwen3.8-2.4T-A95B, a 2.4T-parameter MoE) — under a different, custom 'Qwen3.8-Max License', not Apache 2.0. That license is broadly permissive (use, modify, sell, host, fine-tune) but adds two conditions verified directly from the LICENSE file: if used in a product with >100M monthly active users or >$20M/month revenue, the model name must be displayed; and if you or your affiliates run a Model-as-a-Service or 'AI Work Assistant' business (coding/office-productivity assistants specifically) with combined revenue over $50M in any 12-month period, you need a separate commercial license before using the weights or derivatives commercially. Purely internal use is exempt. South China Morning Post covered this as a deliberate strategy shift away from Alibaba's prior practice of keeping Max-tier models fully closed.

Downloading Qwen3.8-2.4T-A95B gets you a text-only model with forced 'thinking' mode and a native ~262K context window. The hosted qwen3.8-max API additionally supports vision/image input, optional non-thinking mode, and a 1M-token context by default. Alibaba's own model card discloses this distinction, and it drew visible pushback in the model's Hugging Face discussion thread, where a user with the top-voted comment (1.01k+ likes on the repo, active discussion with 36+ replies) argued the release strips 'half its core value' by withholding vision and full context from the open weights while competitors (they cite Kimi K3) shipped full capability.

Pricing is public and tiered by context length. qwen3.8-max (current flagship): $2 / $6 per million input/output tokens (int'l rates, up to 1M context). qwen3.7-max: list $2.5/$7.5, currently 50% off. qwen3.7-plus: $0.4/$1.6 up to 256K context, rising to $1.2/$4.8 beyond that. qwen3.7-flash: as low as $0.03/$0.13 per million tokens for short contexts, scaling to $0.2/$0.8 at 1M context. SCMP reported the qwen3.8-max rates work out to roughly 40% of Claude Opus 5's input price and 24% of its output price at list. New Model Studio accounts also get a free quota (up to 1M tokens per model, valid 90 days) per the official docs.

Artificial Analysis, which runs models through its own held-out suite rather than reporting vendor numbers, scores Qwen3-Max-Thinking at 40 on its Intelligence Index (up from 32 for the preview version), with a Humanity's Last Exam score of 26% — ahead of DeepSeek V3.2 (22%), GLM-4.7 (25%), and MiniMax-M2.1 (22%) on that specific metric, per Artificial Analysis's own write-up. Qwen3-Coder's headline claim of 69.6% on SWE-bench Verified comes from Alibaba's own blog; treat it as a vendor figure pending independent reproduction, though multiple third parties (Nebius, HuggingFace discussion threads) have published their own trajectories against it without disputing the ballpark.

qwen-code (27,078 GitHub stars, Apache 2.0, MIT-based fork lineage) began in 2025 as a close port of Google's Gemini CLI — an early InfoWorld review (Sep 2025) noted it was 'so light an adaptation the documentation still said Gemini.' The project says it stopped syncing with upstream Gemini CLI after v0.1 and has since built out its own feature set (subagents, hooks, MCP, IDE plugins, a desktop app, IM bot integrations). It's now genuinely active: pull requests were merged same-day as this check (16 Aug 2026), and it carries 1,045 open issues, consistent with a project under heavy use rather than dormant. Its README includes a feature-parity comparison table against Claude Code — that comparison is Qwen's own framing, not an independent benchmark, and should be read as a vendor claim.

Independent research has repeatedly found that Qwen models apply Chinese-government-aligned refusals or distortions on a specific set of politically sensitive topics — Tiananmen Square, Xinjiang/Uyghur detention, Taiwan's political status, and related subjects. A 2024 analysis (Shisa.AI, published on Hugging Face) tested Qwen2 across ~95 questions and found refusal rates dropped over 80% when the same questions were asked in Chinese versus English, and that on some topics the model gave confidently wrong answers rather than refusing outright. More recent (2026) mechanistic-interpretability work reports the same refusal pattern persists in Qwen3.5, describing it as a narrow, isolable 'circuit' in the model rather than a broad capability gap — i.e. the underlying knowledge is present but suppressed by alignment training on specific topics. An InfoWorld reviewer testing Qwen Code separately noted the same behavior applies to Qwen's general knowledge answers, though called it irrelevant to code-generation use.