This repository provides FP8-quantized weights for Qwen3.8, a post-trained model compatible with vLLM and SGLang. The official Qwen3.8-Max version adds vision input, 1M context, and built-in tools, offered via Qwen Cloud.\n\nQwen3.8 is the most capable open model from Qwen to date, built on the Qwen3.5 architecture. It features 2.4T total parameters with 95B activated, 92 layers, and a Mixture of Experts with 512 experts, natively supporting 262,144 tokens context (extensible to over 1M).

The model delivers substantial gains in coding, professional work, research, and long-horizon agentic tasks.\n\nBenchmarks show Qwen3.8-Max outperforms previous versions and competitors on SWE-bench Pro (67.7%), Terminal Bench 2.1 (86.6%), and other coding and agentic evaluations, with fine-grained FP8 quantization preserving near-original performance.