Qwen 3.6 27B FP8🔒Private

• 🔒 Runs in a TEE with verifiable hardware attestation
• 🆕 Dense member of Alibaba's Qwen3.6 family
• 📏 Context window around 256K tokens
• 🔧 FP8 quantization roughly halves the weight footprint
• 🧠 Reasoning, agentic coding, and function-calling focused
• 🌐 Venice build surfaces web-search capability
• 📚 Apache 2.0 license permits commercial use and fine-tuning

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,…

Qwen 3.6 27B FP8 is Venice's confidential-computing deployment of Alibaba's Qwen3.6-27B, a dense text model in the Qwen3.6 family.[1] It runs inside a Trusted Execution Environment, exposing hardware attestation evidence so users can independently verify enclave identity and configuration. The Venice build advertises reasoning, code-optimized generation, function-calling, and web-search capabilities, with a context window of roughly 256K tokens. The model is distributed under Apache 2.0, permitting commercial use, fine-tuning, and redistribution.[1]

The FP8 checkpoint quantizes the weights to 8-bit floating point, a path Alibaba publishes alongside its BF16 release to reduce the weight footprint for cost-sensitive deployment.[2] This mirrors the FP8 variant Alibaba offered for the preceding Qwen3.5-27B, so the two generations share the same deployment pattern.[8]

Compared with its same-family sibling Qwen 3.6 35B A3B FP8, a mixture-of-experts model with 35B total and roughly 3B active parameters, this 27B checkpoint is a dense architecture — trading MoE sparsity for a single dense design at a smaller total parameter count. Both are offered here in FP8 inside the same TEE-backed serving environment.

For prospective users, the practical distinction is architectural and operational rather than a ranking claim: the dense 27B activates all parameters per token, while the A3B sibling activates only a fraction of its larger expert pool. Alibaba's Hugging Face model cards remain the authoritative reference for the precise configuration and supported features.[1][2]