Alibaba releases Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model

Alibaba released Qwen3.8-Max, a mixture-of-experts model with 2.4 trillion parameters that activates roughly 95 billion of them per token. The model ranks second in Vision Arena with 1,305 points, 13 points below Claude Fable 5 and two points above Claude Opus 4.7 (Thinking). In a Flappy Bird code-generation test it earned a score of 9 out of 10, matching GPT-5.6 Sol, but at a cost of $0.0248 per run, 4.2 times cheaper than competing models.

The architecture relies on a sparse mixture-of-experts design, where a small router selects only the relevant experts for each token instead of invoking the full 2.4 trillion parameters. This yields a price of $2 per million input tokens and $6 per million output tokens, positioning Qwen3.8-Max on the Pareto frontier in Frontend Code Arena alongside Claude-Opus-5, Kimi-K3, GLM-5.2 and DeepSeek-V4-Flash.

In Vision Arena, which evaluates multimodal image-analysis capabilities, Qwen3.8-Max entered directly at second place with 1,305 points. The gap to first place, Claude Fable 5, is 13 points, while Claude Opus 4.7 (Thinking) trails by two points. The benchmark covers captioning, OCR, diagram understanding and entity recognition.

The Flappy Bird benchmark illustrates the cost advantage: identical 9-out-of-10 score to GPT-5.6 Sol on gameplay, UI and UX, but at $0.0248 versus competitors that are 4.2 times more expensive. The source does not disclose the exact competitor cost, latency figures, or response-time metrics, nor does it provide details on commercial availability, API access or usage limits.

The announcement includes complete documentation on GitHub, from an empty repository to a working application, but without step-by-step guidance. The model appears on two leaderboards, Vision Arena and Frontend Code Arena, and sits on the current Pareto frontier in both. Missing from the release are training resource consumption, market valuation, fundraising information, commercial availability and large-scale production pricing.