Qwen3.6 35B A3B FP8

Dre Dyson's hands-on deployment guide focuses on serving Qwen3.6-35B-A3B-FP8 via vLLM, comparing configurations against a Qwen3.5 baseline. The author rebuilt vLLM from dev wheels (0.19.1rc1) using FlashInfer as the attention backend and the Qwen/Qwen3.6-35B-A3B-FP8 checkpoint on a dual-NVIDIA-GPU rig. Peak throughput The article highlights two model-level improvements called out in the Qwen3.6 release notes: agentic coding enhancements for frontend workflows and repository-level reasoning, and a thinking-preservation feature that retains reasoning context from prior conversation turns. It emphasizes the importance of running an est

This third-party blog from dredyson.com is a practitioner's hands-on retrospective on running Qwen3.6-35B-A3B-FP8, the same model served via Blue Claw's API. The author describes testing on NVIDIA DGX Spark hardware and outlines the three release features that matter to developers: improved agentic coding for frontend While anecdotal rather than authoritative, the article adds useful real-world context for engineers evaluating the Qwen3.6-35B-A3B-FP8 deployment on Blue Claw, especially for agentic coding pipelines where iterative context retention matters. It is not an official Qwen release note or Blue Claw changelog, so feature cl

Millstone AI's inference benchmark page documents Qwen3.6-35B-A3B-FP8 as a 35B-parameter Mixture-of-Experts causal language model with 3B activated parameters per token, organized as 256 experts (8 routed plus 1 shared active per forward pass). The page confirms the FP8 quantization uses fine-grained block-128 quantiza The page highlights agentic coding, repository-level reasoning, and frontend workflow improvements as core focus areas, alongside a toggleable thinking mode and a thinking-preservation feature that retains reasoning traces across conversation turns. Reported hardware benchmarks include 256 tok/s peak throughput on a si