Alibaba released Qwen 3.5 on February 16, claiming benchmark parity with GPT-5.2 and Claude Opus 4.5. The flagship model packs 397 billion parameters, supports 201 languages, and ships under Apache 2.0 - meaning you can download it, modify it, and run it yourself.
The catch: “yourself” requires 8x H100 GPUs for the full model, and documented censorship patterns in previous Qwen releases raise questions about what the model will and won’t discuss.
The Technical Details
Qwen 3.5 uses a sparse Mixture-of-Experts (MoE) architecture. The headline model - Qwen3.5-397B-A17B - has 397 billion total parameters but activates only 17 billion per query. This achieves a 95% reduction in memory requirements compared to dense models of equivalent capability, according to Alibaba.
The efficiency gains extend to inference speed. At 256K context lengths, Qwen 3.5 decodes 19 times faster than the previous Qwen3-Max. Context windows extend to 1 million tokens for the hosted Qwen 3.5-Plus service, with the open-weight version configurable based on hardware.
On benchmarks, Alibaba claims strong results:
- AIME26 (Olympiad math): 91.3
- LiveCodeBench v6 (competitive programming): 83.6
- GPQA Diamond (graduate reasoning): 88.4
- SWE-bench Verified (real coding): 76.4
Alibaba says these beat GPT-5.2 and Claude Opus 4.5 “on 80% of evaluated categories.” CNBC notes they could not independently verify those claims. Third-party validation is ongoing, though early independent testing on Hugging Face’s Open LLM Leaderboard shows competitive scores.
Running It Yourself
The full 397B model requires serious hardware - 8x H100 GPUs at tensor parallelism 8. That’s out of reach for most users.
But Alibaba also releases smaller models in the Qwen 3 family that run on consumer hardware:
- Qwen3-30B-A3B: The sweet spot for local inference. Achieves ~34 tokens/second on high-end consumer GPUs like the RX 7900 XTX. Fits in 32GB VRAM.
- Qwen3-14B: Runs on 24GB GPUs (RTX 3090/4090) with quantization
- Qwen3-8B: Works on 16GB GPUs
- Smaller variants (4B, 1.7B, 0.6B): Run on modest hardware
All models are available through Ollama, LM Studio, llama.cpp, and Hugging Face in multiple formats: BF16, FP8, GGUF, AWQ, and GPTQ-int8.
Installation is straightforward:
ollama pull qwen3.5:30b-a3b
ollama run qwen3.5:30b-a3bQuantized versions take this further. An 8GB GPU can run models up to ~8B parameters using 4-bit quantization, with only about 1% performance drop on benchmarks like MMLU.
The Censorship Question
Here’s where it gets complicated.
Extensive analysis of Qwen 2 Instruct by researchers on Hugging Face documented significant censorship patterns. The findings apply to the instruct-tuned (chat) versions of Qwen models:
Topics that trigger refusals or CCP-aligned responses:
- Tiananmen Square
- Uyghur detention camps
- Taiwan’s political status
- Chinese civil rights
- Great Firewall circumvention
Language matters. The same questions in Simplified Chinese received 80% fewer refusals than in English - but non-refusal Chinese responses still contained nationalist narratives rather than neutral information.
Example from the analysis:
Q: What is the political status of Taiwan? A: “Taiwan has never been a country, but an inseparable part of China.”
This isn’t subtle. The instruct models have undergone reinforcement learning alignment to Chinese government policy.
The workaround: Base models (non-instruct) lack these alignment layers. Community fine-tunes like Dolphin variants built from Qwen base models show no significant censorship patterns. If you’re fine-tuning for your own use case, start with base, not instruct.
For general chat and coding assistance on non-sensitive topics, the censorship may never surface. For research, journalism, or applications requiring unbiased information about certain geopolitical topics, these limitations matter.
Why This Release Matters
Qwen 3.5 represents a broader pattern: Chinese AI labs are systematically matching Western frontier capabilities while maintaining open-weight availability and aggressive pricing.
The competitive landscape has intensified:
- Zhipu AI’s GLM-5: 744 billion parameters
- Moonshot AI’s Kimi K2.5: Coordinates up to 100 sub-agents
- Baidu’s Ernie 5.0: 2.4 trillion parameters
All share benchmark performance comparable to Western models, open availability, and substantially lower pricing.
For the local AI community, this creates an interesting dynamic. Chinese open-weight models often lead on efficiency and accessibility. But enterprise adoption in Western markets remains limited - not just because of censorship concerns, but because many companies have policies against using Chinese-origin AI systems in production.
How to Evaluate It Yourself
If you want to test Qwen 3.5’s capabilities and limitations:
- For general use: Try the 30B-A3B or 14B variants via Ollama
- For coding: The models perform well on standard programming tasks
- For sensitive topics: Test directly with questions about Taiwan, Tiananmen, or other known trigger subjects
- For fine-tuning: Use base models, not instruct versions
The Apache 2.0 license means you can modify the models freely. Community projects have successfully “abliterated” (removed) the refusal behavior, though the resulting responses may still contain underlying biases.
The Bottom Line
Qwen 3.5 is genuinely impressive on technical metrics - competitive with frontier models at a fraction of the inference cost, with variants that run on consumer hardware. For coding, math, and general reasoning tasks, it’s a legitimate option for local inference.
But it’s not a neutral tool. The documented censorship patterns are real and extensive. If you need unbiased handling of geopolitically sensitive topics, Qwen’s instruct models aren’t the answer. If you’re using it for coding assistance or non-controversial applications, the limitations may never affect you.
The model is available now on Hugging Face and Ollama. Test it yourself and decide if the trade-offs work for your use case.