Alibaba Qwen Releases Qwen3.8-Max: A 2.4 Trillion Parameter MoE Model and the Most Capable One in the Qwen Family to Date
Alibaba Qwen team released Qwen3.8-Max, a 2.4 trillion-parameter MoE model with 1M-token context, now generally available via API with open weights coming next week A smaller Qwen3.8-27B checkpoint is also going open-weights, designed for on-premise GPU deployment The model excels in multimodal and agentic capabilities, leading PaperBench (93.0) and IFBench (82.8), with strong vision benchmarks including OSWorld-Verified (86.1) and OmniDocBench 1.5 (92.1) Pricing is $2.00 per 1M input tokens, $6
Analysis
TL;DR
- Alibaba Qwen team released Qwen3.8-Max, a 2.4 trillion-parameter MoE model with 1M-token context, now generally available via API with open weights coming next week
- A smaller Qwen3.8-27B checkpoint is also going open-weights, designed for on-premise GPU deployment
- The model excels in multimodal and agentic capabilities, leading PaperBench (93.0) and IFBench (82.8), with strong vision benchmarks including OSWorld-Verified (86.1) and OmniDocBench 1.5 (92.1)
- Pricing is $2.00 per 1M input tokens, $6.00 per 1M output tokens, with cached input at $0.25 per 1M tokens (8x cheaper than fresh input)
- Major performance gains over predecessor: DeepSWE 1.1 jumped from 21.6 to 56.6, FrontierSWE from 40.7 to 73.5, and JobBench from 31.3 to 53.4
Why It Matters
This release represents a significant step in the open-weight MoE model race, offering both a flagship 2.4T model and a deployable 27B variant that bridges the gap between research and production. The emphasis on multimodal and agentic capabilities over pure reasoning aligns with industry trends toward practical, tool-using AI systems. The pricing structure with aggressive cache discounts provides a cost-effective pathway for enterprises handling repetitive prompt patterns.
Technical Details
- Architecture: 2.4 trillion-parameter mixture-of-experts model accepting text, image, and video input with text output; activated parameter count undisclosed
- Context & Limits: 1M-token context window (991K max input, 983K with thinking enabled), 131K max output, 262K reasoning budget, 2M tokens/minute rate limit
- API Capabilities: OpenAI- and DashScope-compatible, supports function calling, structured outputs, batches, prefix completion, fine-tuning, and five built-in tools (code_interpreter, web_search, web_extractor, t2i_search, i2i_search)
- Benchmark Performance: Terminal-Bench 2.1 at 86.6 (behind GPT-5.6 Sol's 88.8, ahead of Claude Opus 4.8/Fable 5 at 84.6); SWE-bench Pro at 67.7; GPQA Diamond at 92.6; strong vision scores across OSWorld-Verified, Parametric CAD Bench, and OmniDocBench
- Deployment Options: Hosted API available immediately; open weights for both Qwen3.8-Max and Qwen3.8-27B shipping next week; 27B checkpoint targets on-premise GPU hardware
Industry Insight
The dual-release strategy (flagship + compact) mirrors a growing industry pattern where companies cater to both cloud-native and on-premise deployment needs, lowering barriers for enterprises with data sovereignty requirements. The 8x cost premium for uncached versus cached input suggests that prompt engineering and prefix optimization should be prioritized in production deployments to control expenses. The benchmark caveats—particularly comparing multimodal results against Qwen3.7-Plus rather than Qwen3.7-Max—highlight the importance of scrutinizing release claims and maintaining independent evaluation standards.
Disclaimer: The above content is generated by AI and is for reference only.