Qwen3.6 35B A3B Uncensored🔒Private

• - 🧠 Mixture-of-experts: 35B total parameters, ~3B active per token
• - 🆕 Community "uncensored" fine-tune of Alibaba's Qwen3.6-35B-A3B base
• - 🔒 Runs in a Trusted Execution Environment with hardware attestation
• - 📏 Native context up to 262,144 tokens
• - 👁️ Catalog lists multimodal input across text, images, and video
• - 🔧 Vendor reports big agentic-coding gain over Qwen3.5-35B-A3B
• - 🌐 Capabilities include web search; served at FP8 quantization
• - 🆕 Adds "Thinking Preservation" to retain reasoning across turns

Alibaba

Alibaba Group is a Chinese multinational technology company founded in 1999 and headquartered in Hangzhou, Zhejiang. Originally built around e-commerce and cloud computing, Alibaba has become one of the most prolific contributors to open-weight AI research,…

Qwen3.6 35B A3B Uncensored is a community-modified variant of Alibaba's Qwen3.6-35B-A3B, a sparse mixture-of-experts model with 35 billion total and roughly 3 billion active parameters. The "uncensored" descriptor refers to a third-party post-training fine-tune that strips the base model's safety refusal mechanisms without altering core capabilities; it is not an official Alibaba release. Venice runs it inside a Trusted Execution Environment with attestation evidence for independent verification, here at FP8 quantization.

Architecturally, the base uses 256 experts routing 8 per token, 40 layers, and a hybrid linear plus full-softmax attention mechanism, supporting native context up to 262,144 tokens. The catalog additionally lists multimodal input across text, images, and video. Compared with its same-family predecessor Qwen 3.5 35B A3B from February 2026, the Qwen3.6 generation adds the "Thinking Preservation" option to retain reasoning context across messages. Per Alibaba's own blog, Qwen3.6-35B-A3B "surpasses its predecessor Qwen3.5-35B-A3B by a wide margin" on agentic coding.

This entry is the uncensored sibling of the standard Qwen 3.6 35B A3B FP8 deployment, sharing the same base weights and TEE delivery but removing refusal behavior. Both maintain Qwen3.6's thinking and non-thinking modes. Note that the catalog configures a 128K-token context window for this deployment, below the architecture's native 262K maximum.