Alibaba's Qwen team just shipped its biggest model yet. Qwen3.8-Max, made generally available on August 3, 2026, is a 2.4-trillion-parameter mixture-of-experts (MoE) model that accepts text, images, and video and returns text. It is the most capable model in the Qwen family to date, and the team says open weights are coming next week.
Here is what actually matters, and where the fine print deserves a second look.
A frontier-scale MoE with a 1M-token window
At 2.4 trillion total parameters, Qwen3.8-Max sits firmly in frontier territory on paper. It is a sparse MoE, so only a fraction of those parameters fire on any given token. Alibaba has not officially confirmed the activated-parameter count — third-party trackers estimate roughly 95B active per token — which means you cannot yet model serving cost precisely from the architecture alone.
The context window is the headline number for long-context workloads:
| Spec | Value |
|---|---|
| Total parameters | 2.4T (MoE) |
| Context window | 1,000,000 tokens |
| Max input | 991K tokens (983K with thinking on) |
| Max output | 131K tokens |
| Max reasoning budget | 262K tokens |
| Rate limits | 2M tokens/min, 15K requests/min |
The model is OpenAI- and DashScope-compatible, so wiring it into an existing stack is a base-URL and model-ID change rather than a rewrite. It supports function calling, structured outputs, batches, prefix completion, and fine-tuning. Five built-in tools ship on the Responses API: code_interpreter, web_search, web_extractor, t2i_search, and i2i_search.
Pricing: cheap for a flagship, with a cache twist
Qwen3.8-Max is priced aggressively for a model this size:
- $2.00 per 1M input tokens
- $6.00 per 1M output tokens
- $0.25 per 1M tokens for implicit cache reads
Explicit caching costs $2.50 to create and $0.17 per 1M tokens to read. That makes cached input roughly eight times cheaper than fresh input, which flips the usual optimization advice on its head: with Qwen3.8-Max, prefix stability drives your bill more than raw prompt length does. If you are running repository-scale coding agents or long-document knowledge bases, structure your prompts so the stable prefix stays identical across calls.
For comparison, that $2/$6 rate undercuts most Western frontier APIs while offering the full 1M-token window at a single flat tier.
The benchmarks — and the asterisks
Alibaba published a full benchmark table with the release. The gains are real, but read them with the caveat that every number comes from Alibaba's own internal evaluations, not from third parties like Artificial Analysis or LMArena.
Qwen3.8-Max leads PaperBench at 93.0 and IFBench at 82.8, and posts GPQA Diamond at 92.6 — a marginal bump over Qwen3.7-Max's 92.4.
On agentic and coding work the picture is more mixed. It scores 86.6 on Terminal-Bench 2.1, ahead of Claude Opus 4.8 and Claude Fable 5 (both 84.6) but behind GPT-5.6 Sol at 88.8. On SWE-bench Pro it reaches 67.7, trailing Fable 5's 80.0, and on FrontierSWE it lands 73.5 against Fable 5's 88.8. So while it is competitive, it does not top the frontier for pure software engineering.
Where it clearly shines is multimodal and agentic tasks: it tops most vision rows, including OSWorld-Verified at 86.1, Parametric CAD Bench at 91.5, and OmniDocBench 1.5 at 92.1.
Two honest caveats belong in any read of that table. First, the multimodal comparisons are against Qwen3.7-Plus, not Qwen3.7-Max, which flatters the generational jump. Second, Alibaba's own RL scaling curve peaks at 0.725 near 4,000 training environments, then declines to 0.719 and 0.689 — a candid signal that the team is hitting diminishing returns on reinforcement scaling.
Can you actually run it?
That depends on which artifact you mean. The hosted API is deployable today by any team, and given the OpenAI-compatible surface, integration is trivial.
The open weights are a different story. At 2.4T total parameters, the flagship checkpoint is a multi-node datacenter artifact — not something you spin up on a workstation. The realistic self-hosting path is the smaller Qwen3.8-27B checkpoint, also going open-weights, which fits ordinary on-premise GPU hardware. If your goal is local deployment, watch for the 27B, not the flagship.
The Bottom Line
Qwen3.8-Max is a serious flagship: a 2.4T-parameter MoE with a genuine 1M-token window, strong multimodal and agentic scores, and pricing that undercuts most Western rivals at $2/$6 per million tokens. The open-weights promise — for both the flagship and the deployable 27B — could make it a default for cost-sensitive long-context and vision workloads.
Just keep the asterisks in view: the benchmarks are self-reported, the activated-parameter count is undisclosed, and the software-engineering numbers still trail the very best. For most teams, the smart move is to run it against your own tasks through the API this week, and revisit the on-premise question when the 27B weights land.
