Alibaba releases Qwen3.8-Max, a 2.4 trillion parameter multimodal AI model with coding and autonomous agent capabilities

Decision brief

What changed

Alibaba released Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts multimodal model (text, image, video) available via a hosted API at $2 per million input tokens ($6 per million output), with open weights slated for Hugging Face and ModelScope; a smaller 27B checkpoint is offered for on-premise use. Alibaba claims the model matches top OpenAI and Anthropic frontier models and demonstrated a 16-day autonomous coding run with 265 public GitHub commits.

Why it matters

This is another concrete data point in the US-China frontier AI race, offering a lower-cost, partially open alternative to US frontier APIs that could affect vendor selection, cost structures, and build-vs-buy decisions for AI-heavy workloads. The open-weights option changes leverage in vendor negotiations, but the full 2.4T model's memory footprint makes true self-hosting impractical for most organizations, so most buyers will still depend on Alibaba's hosted infrastructure with attendant data residency and security considerations.

Affected roles

CEO CTO CISO CFO COO

Evidence

Three independent outlets (The Verge, MarkTechPost, The New Stack) consistently report the same core specs—2.4T parameters, MoE architecture, 95B active parameters per token, $2/million input token pricing, and planned open-weight release—giving reasonable confidence in the technical facts; however, performance-parity claims against OpenAI/Anthropic originate from Alibaba itself and are only relayed, not independently verified, by the coverage.

What remains uncertain

Claims that Qwen3.8-Max matches or nearly matches top US models (including a reference to Anthropic's 'Fable 5,' an unusual model name that may reflect a reporting inconsistency) are self-reported and unbenchmarked by third parties. The quality, security, and reliability of the 16-day autonomous coding demonstration are also unverified beyond the commit count, and the exact timing/completeness of the open-weights release is not yet confirmed.

Monitor next

Watch for independent third-party benchmark results (e.g., Chatbot Arena, LMSYS, or enterprise evaluations) once open weights are actually published on Hugging Face and ModelScope.

Analytical support, not advice — assumptions and open questions stated above.

Source coverage