2.4T total params, ~95B active per token (MoE). Built on the Qwen 3.5 architectural foundation with a hybrid attention mechanism; roughly 4% expert activation per forward pass. Native 1M context, 131K max output, 262K max reasoning. Multimodal: text + image in, text out. Reasoning controls: reasoning_effort (xhigh default / medium / low), enable_thinking, preserve_thinking.
Launched 2026-08-03 on Alibaba Cloud Model Studio (DashScope) and QwenWork - the first Max-class Qwen to get a public per-token API. List price $2/M input (flat across the full 1M context), $6/M output, $0.25/M implicit cache hit - cheaper than the Qwen3.7-Max list price ($2.50/$7.50). OpenAI- and Anthropic-protocol compatible; three regional endpoints (Beijing, Singapore, US-Virginia). First shown 2026-07-19 at WAIC Shanghai.
Open weights promised ~2026-08-10 on HuggingFace and ModelScope (Apache-2.0) - the first open-weight Max-class Qwen ever. Until the checkpoint ships, no model_variants are seeded (quant sizes unknown) and the model is cloud-only, so it does not appear on the local-run pages (e.g. /dgx-spark) and the “Run it locally” / “Download options” sections stay off the model card.
Vendor benchmarks: GPQA-Diamond 92.6, PaperBench 93.0, IFBench 82.8, MathVision 95.2, LogicVista 91.9, OSWorld-Verified 86.1. Weakest flagship scores: HLE 43.6, SWE-bench Pro 67.7, FrontierSWE 73.5 (Fable 5 leads both at 80.0 / 88.8). Arena: 5th Text, 2nd Vision, 4th Frontend Code. Treat the agentic-coding figures as vendor-reported pending independent replication.
AI-generated content marks
This model embeds text watermarks in generated text and adds C2PA provenance metadata to supported files such as .png, .jpg, and .svg.
Marks can be lost through editing, screenshots, or format conversion, so their absence does not prove a file is human-made.
Provider transparency docs →