Alibaba Qwen3.8-Max Preview: 2.4 Trillion Parameters, No Benchmarks, Open-Weight Promise

Context window and input modalities

The model inherits a 1 million-token context window from its predecessor Qwen3.7-Max . It natively processes text, images, video, and documents in a single model, rather than relying on a separate vision variant as earlier Qwen models did . This native multimodality at scale is a significant technical step forward for the Qwen family.

Performance claims

Alibaba states the model ranks "second only to Anthropic's Claude Fable 5" among the systems it internally benchmarked . The company also says it beats Qwen3.7-Max on coding, full-stack development, data analysis, agentic workflows, and long-horizon "professional cowork" tasks . No external benchmark scores, model card, or technical report were published alongside the announcement . Independent reviewers and platforms such as LMArena and Artificial Analysis have not yet listed the model . One independent review notes that the endpoint responds, but every benchmark number circulating "for" Qwen3.8 is actually from Qwen3.7-Max testing .

Open-weight plans

Alibaba stated the model will go open-weight "soon" . As of late July 2026, no specific date, license, weight-download link, or Hugging Face repository had been published . The open-weight promise is a major strategic signal — if fulfilled, Qwen3.8-Max would be the largest openly available model to date — but remains unfulfilled.

Pricing and access

The preview is available through three channels :

• Alibaba's Token Plan — subscription tiers roughly $6, $18, and $68 per month, with Alibaba offering the model at 10% of standard price as a launch promotion .
• Qoder and QoderWork — Alibaba's coding platforms.

Standard pay-as-you-go API pricing per million input/output tokens has not been published . The endpoints are compatible with both the OpenAI and Anthropic API protocols, meaning developers can drop the model into existing tooling like Claude Code, Cursor, and OpenCode without changing their harness .

Broader context: China's AI race intensifies

Qwen3.8-Max arrived just days after Moonshot AI released Kimi K3 as open-weight, signaling an acceleration in the already intense Chinese AI competition . The Chinese market in 2026 has seen a wave of aggressive flagship launches and open-weight pledges from both established tech giants (Alibaba, Baidu, Tencent) and well-funded startups (Moonshot, Zhipu AI, 01.AI, DeepSeek) .

Alibaba's decision to preview a 2.4-trillion-parameter model while publishing zero benchmarks underscores a broader pattern: Chinese AI labs are racing to claim scale leadership with large parameter counts and "open-weight" commitments, often before independent validation is possible . The Qwen3.8-Max announcement is as much a strategic move in this competitive landscape as it is a technical release.

The bottom line

Qwen3.8-Max-Preview is a real, testable API endpoint with impressive nominal specifications — 2.4 trillion parameters, native multimodality, million-token context — and an ambitious but unverified performance claim. For developers, the preview offers a chance to experiment through subscription access. For the AI community, the model's true standing will remain unclear until Alibaba publishes benchmarks, a model card, and open weights. Until then, the 2.4 trillion-parameter count and "second only to Fable 5" ranking remain the company's self-assessment, not an independently verified fact .