Qwen3.8-Max: Alibaba's 2.4T Open Model Hits the Arena

Alibaba Cloud’s Qwen team released Qwen3.8-Max on August 3 — a sparse mixture-of-experts model with 2.4 trillion total parameters of which only 95 billion are active per token, a 1-million-token context window, and open weights. The team detailed the architecture on its official blog, and the developer weights are now public on Hugging Face and ModelScope.

Key facts

• 2.4T total / 95B active per token: the sparse architecture keeps inference costs a fraction of the dense model of comparable size.
• 1-million-token context: built for long-horizon work — deep code refactors, agentic workflows, document-scale analysis.
• Leaderboards on release: #2 on the Vision Arena, top-five text on Text Arena — the strongest launch position this quarter for an open-weight model.
• Open weights now live on Hugging Face and ModelScope, continuing the Qwen family’s publishing tradition.
• API pricing: roughly ₹2.00 per-million input tokens and ₹6.00 per-million output tokens — compared to GPT-5.6 Sol Standard in the upstream API. (See our price-cuts briefing for the frontier pricing table.)

The announcement

Alibaba positioned Qwen3.8-Max as an answer to two gaps in the market: open-weight models that are genuinely competitive at the frontier, and APIs priced for volume agentic work. The announcement emphasizes the model’s design for long-horizon, context-heavy workloads, and confirms the weights are free to download, run, and fine-tune.

Because the model’s architecture activates only 95B parameters per token, serving costs scale with active — not total — parameters. That is the engineering trick that lets a 2.4T model charge less than many 400B-class models.

Why it matters

• Long context without a data-center budget. Teams that need 1M-token context previously had two choices: pricey closed tiers or clunky chunking. Qwen3.8-Max open-weights make the whole document the prompt at close to commodity cost.
• Vision + sparse at scale. Second place on the Vision Arena puts the open stack within reach of the closed frontier on multimodal work for the first time this quarter.
• Another round in the price war. Combined with the open weights model cost, the API price pressures closed vendors further — especially on input tokens where agent workloads live.
• An ecosystem choice. The Qwen family is already the most widely used open-weight series among developers this year; a frontier-grade release in the same family strengthens that funnel further.

What to watch

• Arena movement. Leaderboard positions skew early-vote; watch the Text Arena and Vision Arena positions as the voting base grows over the next weeks.
• Fine-tune adoption. Which agent frameworks and tooling ecosystems add Qwen3.8-Max adapters first — a leading indicator of developer staying power.
• The Qwen cadence. The Alibaba team has shipped a generational frontier release every few months; the existence of “Qwen3.8” as a jump-milestone implies active roadmap.

Open source

• Qwen blog: Qwen3.8-Max announcement
• Weights: Hugging Face (Qwen/Qwen3.8-Max) and ModelScope
• API pricing: Alibaba Cloud’s Model Studio
• Shared-basis comparison: Qwen3.8-Max vs Kimi K3: The Numbers Beat

Updated August 8, 2026 — weights now available.