REVIEWS / AI MODELS / OWNER INSIGHTS
🦉 WE READ 333 OWNER COMMENTS
Qwen 3: what owners actually say
Owners praise Qwen 3's genuine open-source release and strong local performance, but censorship on China-related topics and thinking-model token bloat are persistent frustrations.
What owners complain about
- China-related censorship and propaganda COMMON
Multiple users report the model refuses or sanitizes answers about Taiwan ('Taiwan is an inalienable part of China') and Tiananmen Square. One user observed it 'definitely knows about Tiananmen' but 'gets shot in the head the second it decides to talk about it.' Censored vs uncensored versions behave differently.
- Thinking models waste tokens and feel slow SOME
Users report thinking variants output ~1000+ tokens of internal reasoning before responding, making even smaller MoE models (35-A3B) feel slower than larger non-thinking ones (80-A3B). Described as 'wasting sooooo much time and so many tokens.'
- High variance in Coder output quality SOME
Qwen Coder 'sometimes can match top of the line and other times makes basic mistakes.' Users note it needs careful steering to perform well, and many are still waiting for a 3.6 Coder release that doesn't yet exist.
- De facto prohibition in government contracting FEW
All DeepSeek and Qwen models are perceived as prohibited in US government contracting, including local deployments via Ollama — 'no legislative or executive mandate yet exists, but it's perceived as a gap.'
- Quantization and hardware complexity SOME
Getting good local performance requires deep understanding of quantization methods (Q4_K, K_XL, etc.) and careful hardware trade-offs. Users debate 3090s, Intel Arc B70s, and Mac M1-M4 bandwidth, with one noting Vulkan mode runs 'really slow' despite good VRAM.
What owners love
- Genuinely open source — not just open weights
Users highlight that Qwen releases model weights, training hyperparameters, datasets, AND code — making it one of the very few models compliant with the Open Software Institute's definition of open source AI, unlike Llama, Gemma, and other big-tech 'open' models.
- Strong local coding and agentic performance
Owners report the Coder variant is 'fast as f***' and competitive with top-line models for local tasks. One user noted improvement over previous versions in 'discovering new creative pathways' in agentic wiki-exploration tasks. Benchmarks land between Gemma 4 26B and Qwen 3.6 35B for coding.
- Open weights drive down inference costs dramatically
Cerebras running Qwen 3 235B Instruct at 1.4k tokens/second 'for cheaper than Claude Haiku' — users emphasize open weights put inference on the open market, benefiting all providers and users.
- Practical reasoning quality on edge cases
One user tested the 'car wash puzzle' (50m distance — walk or drive?) and the model correctly reasoned it should walk, explaining the negligible distance makes driving wasteful.
Surprising patterns
- Censorship is platform-dependent: the same model in HuggingChat shows fewer restrictions than the official demo — suggesting filtering is applied at the demo layer, not the model itself.
- Users report the model has detailed knowledge of censored topics (e.g., Tiananmen Square events) and begins to answer before being interrupted, rather than lacking the knowledge entirely.
- A larger MoE model (80-A3B) can feel 'massively faster' than a smaller one (35-A3B) simply because the smaller variant is a thinking model — the token overhead of reasoning outweighs the compute savings from fewer active parameters.
WHO SHOULD SKIP IT
US government contractors and anyone needing reliable, uncensored answers on topics sensitive to the Chinese government should skip Qwen 3, as censorship is deeply embedded and regulatory acceptance is uncertain.
Synthesised from 333 real owner comments across 6 platforms. Every point is grounded in the comments — no marketing, no AI guessing. How we do it →