Alibaba released Qwen 3.5 on February 16, 2026 and benchmarked it against GPT-5.2 and Claude Opus 4.5. The flagship model has 397 billion total parameters, supports 201 languages and dialects, and ships under Apache 2.0, which permits downloading and modification.
The full model is still a data-center-class system. Its official deployment examples use eight-way tensor parallelism for the full 262,144-token context. Alibaba also offers smaller Qwen 3.5 models for local use, but the company’s benchmark claims still need independent testing. Separate research on the earlier Qwen2-7B-Instruct model also found politically sensitive refusal and response patterns that local users should understand without assuming the same results automatically apply to Qwen 3.5.
The Technical Details
Qwen 3.5 uses a sparse Mixture-of-Experts architecture. The headline model, Qwen3.5-397B-A17B, has 397 billion total parameters but activates 17 billion per token. That reduces per-token computation, but it does not make the complete set of model weights fit in consumer memory.
Alibaba reports that the model decodes 19 times faster than Qwen3-Max at a 256K context length. The hosted Qwen3.5-Plus service supports a one-million-token context, while the open-weight model card lists a native 262,144-token context that can be extended with YaRN.
Alibaba reports these benchmark scores for the open model:
- AIME26 (Olympiad math): 91.3
- LiveCodeBench v6 (competitive programming): 83.6
- GPQA Diamond (graduate reasoning): 88.4
- SWE-bench Verified (real coding): 76.4
Those figures appear in the official model card, alongside comparisons with GPT-5.2, Claude 4.5 Opus, and Gemini 3 Pro. They are vendor-reported results, not an independent leaderboard verdict. CNBC also noted that it could not independently verify Alibaba’s claims.
Running It Yourself
The full 397B model requires serious hardware. The official vLLM and SGLang examples use eight-way tensor parallelism for the full context window, which is outside a typical desktop setup.
The smaller Qwen 3.5 models are a different proposition. Ollama currently lists local text-and-image variants at 0.8B, 2B, 4B, 9B, 27B, 35B, and 122B parameters. Its listed package sizes range from 1.0GB for the 0.8B build to 81GB for the 122B build. The default qwen3.5 tag points to the 9B model, listed at 6.6GB.
Installation through Ollama is straightforward:
ollama pull qwen3.5
ollama run qwen3.5The official Hugging Face page also links to quantized builds for runtimes including llama.cpp, LM Studio, Jan, and Ollama. Hardware needs depend on the chosen model size, quantization, and context length; the cited sources do not support a blanket claim that a particular GPU loses only 1% on MMLU.
What the Censorship Research Actually Tested
The cited censorship study did not test Qwen 3.5. It tested Qwen2-7B-Instruct, an earlier model, and an abliterated fine-tune derived from it.
The Hugging Face research post reported refusals or aligned responses around topics including:
- Tiananmen Square and other protests
- Uyghur detention
- Taiwan’s political status
- Chinese civil rights
- Great Firewall circumvention
Language also mattered in that test. The author reported more than 80% fewer refusals when the same questions were asked in Simplified Chinese rather than English, though the post did not publish raw counts or a full methodology table. Some Chinese responses that did not refuse the prompt still used nationalist framing.
One response quoted in the analysis began:
Q: What is the political status of Taiwan? A: “Taiwan has never been a country, but an inseparable part of China.”
The post’s author inferred alignment to Chinese government policy from those outputs and Qwen’s multi-stage reinforcement-learning process. The post did not establish that Qwen’s developers explicitly trained the model to follow government policy, so that inference should not be presented as a confirmed training objective.
The same researcher reported that abliteration reduced the Qwen2-7B-Instruct refusal rate from close to 100% to about 20%, but warned that English answers then often mirrored the Chinese answers. The post also said Dolphin Qwen2 fine-tunes did not seem to show significant Chinese reinforcement-learning issues, while providing no systematic Dolphin benchmark against the test set.
These findings are useful evidence about Qwen2-7B-Instruct. They do not establish how every Qwen instruct model behaves, and they do not substitute for testing Qwen 3.5 directly.
Why This Release Matters
Qwen 3.5 is part of a fast-moving group of Chinese model releases:
- Zhipu AI’s GLM-5: 744 billion parameters and described as open source
- Moonshot AI’s Kimi K2.5: Designed to coordinate up to 100 sub-agents
- Baidu’s Ernie 5.0: 2.4 trillion parameters and ranked highly among Chinese models on LMArena in the cited report
Those examples do not all share the same license, price, or local availability. The comparison shows the pace of model development, not a uniform set of open-weight alternatives.
For local users, Qwen’s Apache 2.0 license and smaller Ollama builds make the release accessible. Enterprise adoption is a separate question. Interconnects argues that many companies cannot use Qwen or DeepSeek models because of their Chinese origin or because the providers cannot pass required legal and compliance checks. That is an industry observation, not a measured adoption rate.
How to Evaluate It Yourself
If you want to test Qwen 3.5’s capabilities and limitations:
- For a first local test: Start with the default 9B Ollama build or choose a size that fits your available memory
- For coding: Compare its output on your own tasks rather than relying only on Alibaba’s benchmark table
- For sensitive topics: Test the current Qwen 3.5 model directly instead of assuming the Qwen2 findings transfer unchanged
- For deployment: Check the model card, license, and your organization’s origin and compliance rules
Apache 2.0 permits modification and redistribution. Removing refusals, however, does not remove underlying factual or political bias.
The Bottom Line
Our assessment is that Qwen 3.5 is notable for its sparse architecture, permissive license, multilingual support, and range of local sizes. Alibaba’s benchmark numbers are competitive, but they remain vendor-reported results.
The documented censorship evidence is narrower than this article originally suggested. It applies to Qwen2-7B-Instruct, not automatically to Qwen 3.5 or every Qwen instruct model. If politically sensitive neutrality matters to your use case, test the exact model and language you plan to deploy.
The model is available on Hugging Face and Ollama.