Qwen3.8 Max: 2.4T MoE, $2/M Tokens, Open Weights (2026)

Qwen3.8 Max is the flagship model that Alibaba's Qwen team announced on August 3, 2026. It is a sparse mixture-of-experts (MoE) model with 2.4T total parameters, activating roughly 95B parameters per token (an activation ratio of about 4%). It supports up to 1 million tokens of context and up to 128k tokens of output. API pricing starts at $2.00 per million input tokens and $6.00 per million output tokens, with cached input priced at $0.25. Notably, Alibaba has announced that Qwen3.8 Max will be the first Max-class model to have its weights released, and the smaller Qwen3.8-27B will be released as open weights at the same time.

Qwen3.8 Max uses a sparse mixture-of-experts design: of its 2.4T total parameters, only about 95B — roughly 4% — are actually used for computation on any given token. This design appears intended to balance the representational capacity of a large parameter count against the compute cost of inference. It should be noted, however, that even the 95B active parameter figure is far too large for typical local, personal-hardware inference, and running the full 2.4T model would require substantially larger infrastructure still.

All of these figures come from third-party evaluations. The SWE-bench resolve rate of 87.3% is reported to exceed GPT-5.5's 82.6%. On the Vals Index, Qwen3.8 Max is reported to match Claude Opus 4.7's score at roughly 1/2.3 the cost, positioning it as a cost-performance play.

This pricing reflects the figures at announcement; whether there is tiered pricing by context length, or regional price differences, has not been confirmed as of this writing. Some other Qwen-series models do use tiered pricing that rises with context length, so it is worth checking the official documentation before committing to heavy usage. For a look at how tiered pricing works on a related Qwen model, see Qwen3.7 Flash's tiered pricing breakdown.

Qwen3.8 Max is notable as the first Max-class model from Alibaba to have an open-weight release announced, with the release expected "next week." Max-class models have historically been closed and API-only, so this marks a shift in approach. At the same time, the smaller Qwen3.8-27B is also planned for open-weight release. However, its detailed specs — whether it is a dense or MoE architecture, and VRAM requirements at various quantization levels — remain unconfirmed as of this writing, pending an official spec announcement.

- Detailed specs for the 27B model (dense vs. MoE, VRAM by quantization) are unconfirmed
- Local execution of Qwen3.8 Max itself (2.4T) is effectively impractical on personal hardware
- The exact weight-release date is unconfirmed (only "next week" has been stated)
- License terms are unconfirmed

Alibaba's announcement emphasizes long-horizon, autonomous agentic use cases for Qwen3.8 Max. The specific examples cited include:

- Autonomous coding work sustained over 10+ days
- Chip design optimization processes spanning 500+ turns
- Year-long (365-day) e-commerce strategy planning and execution

Qwen3.8 Max leads with strong overall coding performance and a pitch for sustained autonomy in agentic work, with its 87.3% SWE-bench resolve rate positioned above GPT-5.5. For vision-heavy or low-cost batch-processing needs, a lower tier within the same family such as Qwen3.7 Flash may offer a better price point. For agentic coding or local-deployment priorities, Qwen3.6 Plus's agentic coding guide and Qwen 3.6 27B (dense architecture) are also worth comparing. Once weights are released, the newly announced Qwen3.8-27B could become a strong candidate for local deployment, but with its detailed specs still unconfirmed, that judgment should be held for now.