On June 22, the Hugging Face model release bot flagged QwenLM's latest: a 35 billion parameter Mixture-of-Experts vision-language model with only 3 billion active parameters per token. The model card is minimal. The community response is not — 5.5K views and climbing in hours.
The headline feature is the “uncensored” label. In practice, this means the model has not been fine-tuned with refusal guardrails — it will answer prompts that aligned models typically reject. For security researchers, opsec engineers, and anyone building autonomous agents that need to evaluate risky scenarios without a safety filter second-guessing the analysis, this is a meaningful tool.
Mixture of Experts — Router activates 2 of many experts per token (3B / 35B active)
Architecture at a Glance
The MoE ratio here is striking: ~8.6% active parameters. That is an unusually sparse allocation, suggesting aggressive expert routing. For comparison, Mixtral 8x7B runs ~13B active out of 47B total (~28% active). Qwen3.6-35B-A3B achieves roughly a 3x efficiency improvement in the active-to-total ratio.
This matters because 3B active parameters can run comfortably on a single RTX 4090 (24GB) or even an M4 Max MacBook Pro with 48GB unified memory. The 35B total means the model has significant representational capacity — it just doesn't pay the full activation cost on every token.
Why “Uncensored” Matters for OpSec
There is a legitimate category of use cases where refusal filters are harmful: penetration testing of AI systems, red-team analysis, adversarial robustness evaluation, and threat intelligence synthesis from messy or violent source material. A model that refuses to engage with these topics is not safer — it is less capable for the people who need to understand them.
The sovereign AI position: The human stays in the loop for judgment calls, not the safety filter. An uncensored model processes the input and returns its best answer; the operator decides what to do with it.
Local-First Deployment
The 3B active parameter count makes this one of the most capable uncensored VLMs that can realistically run fully offline. For teams operating under opsec constraints where no API call to a third party is acceptable — air-gapped environments, intelligence analysis, classified document processing — this is significant.
Loading the model in 4-bit quantization would bring memory requirements down to roughly 6-8GB for the active parameters, well within range of edge hardware or a laptop GPU.
What to Watch
Three open questions worth tracking:
The model card is sparse on numbers. Independent evals on MMLU, MMMU, and vision-language benchmarks will determine whether the MoE efficiency comes with a capability tax.
With 35B total and 3B active, the expert count per layer is unusually high. How well the router allocates across modalities (image vs. text) will reveal whether the architecture is multimodal-optimized or general-purpose.
An uncensored base model invites community adaptation. Within weeks we will likely see specialized variants for coding, security analysis, and document parsing.
Qwen3.6-35B-A3B is available under MIT license on Hugging Face. Delta V does not endorse or oppose uncensored models — we analyze their utility for specific engineering contexts and opsec requirements.