- Published on
A 2.4 Trillion Parameter MoE Model and the Most Capable One in the Qwen Family to Date
- Authors
- Name
- DeepHorus
Alibaba's Qwen team has released Qwen3.8-Max, its most capable model to date, making it broadly available via API while confirming that open weights for both the flagship and a smaller 27B checkpoint will ship next week. The 2.4-trillion-parameter mixture-of-experts model marks the next step in the Qwen lineage, succeeding Qwen3.7-Max with notable gains in multimodal and agentic tasks. It accepts text, image, and video inputs and returns text, positioning it against leading frontier models from OpenAI, Google, and Anthropic.
The model supports a 1-million-token context window, with maximum input of 991K tokens (983K with thinking enabled) and maximum output of 131K tokens. The reasoning budget extends to 262K tokens. Alibaba has set rate limits at 2 million tokens per minute and 15,000 requests per minute. Access costs $2.00 per million input tokens and $6.00 per million output tokens. Implicit cache reads are priced at $0.25 per million tokens, while explicit cache creation is $2.50 and explicit cache reads $0.17 per million. Cached input is eight times cheaper than fresh input, making prefix stability a key cost driver.
Built on a mixture-of-experts architecture, the model's activated-parameter count has not been disclosed, making on-premises serving costs difficult to model. The hosted API is OpenAI- and DashScope-compatible, requiring only a base-URL and model-ID change to integrate. Five built-in tools ship on the Responses API, including code_interpreter, web_search, web_extractor, t2i_search, and i2i_search. Alibaba published a benchmark table showing strong results. Qwen3.8-Max scores 86.6 on Terminal-Bench 2.1, ahead of Claude Opus 4.8 and Claude Fable 5 at 84.6, though behind GPT-5.6 Sol (max) at 88.8. It leads PaperBench at 93.0 and IFBench at 82.8, and reaches 92.6 on GPQA Diamond. Multimodal gains are dramatic, with OSWorld-Verified 86.1 and OmniDocBench 1.5 at 92.1. However, it trails Claude Fable 5 on SWE-bench Pro (67.7 vs 80.0) and FrontierSWE (73.5 vs 88.8).
The model targets software engineering, legal and financial document review, media and e-commerce operations, and design. Practical applications include repository-scale coding agents, long-document knowledge bases, long-video indexing, structured data extraction, and multi-step research assistants. The Qwen3.8-27B checkpoint is the realistic path for on-premise GPU hardware, while the 2.4T flagship is a multi-node datacenter artifact. Industry observers note that the pricing is aggressive for a frontier model, and the open-weights promise has generated excitement in the open-source community. Analysts caution that the lack of activated-parameter details and license information makes full evaluation difficult. With open weights expected next week, developers will soon be able to verify the benchmarks independently. The release intensifies competition in the multimodal frontier, as Alibaba continues to close the gap with Western labs while offering a more accessible API.
