Alibaba Published Qwen3.8-Max's Benchmarks. Every One of Them Is Alibaba's.

HANGZHOU, 8 AUG 2026 — Alibaba made Qwen3.8-Max generally available on 3 August and, for the first time, published benchmark scores for it. Every one of those scores is Alibaba's own.

When we covered the preview on 22 July, the missing numbers were the story: Alibaba Previews Qwen3.8-Max at 2.4 Trillion Parameters, With No Benchmarks Published. They have arrived. The results cut in both directions: two points ahead of Anthropic's models on one agentic benchmark, twelve points behind Fable 5 on a coding benchmark. The open-weights release, promised for the following week, is the part with consequences beyond a leaderboard.

What shipped on 3 August

The model is a mixture-of-experts design with 2.4 trillion total parameters and a one-million-token context window, split as a maximum 991,000 tokens of input and 131,000 of output. It is served through Alibaba Cloud's Model Studio APIs, which accept both OpenAI-compatible and DashScope calls, and through QwenWork, the company's workplace agent platform.

API pricing is US$2 per million input tokens, US$6 per million output, and US$0.25 per million cached input. For comparison against the rest of the field, our own frontier-model rate comparison lists the current published prices for eleven models. That is a RECATOOLS page, and we maintain it ourselves.

Alibaba also said weights for both Qwen3.8-Max and a smaller Qwen3.8-27B would follow within the week, and recommended the 27B for on-premise GPU hardware.

That asymmetry in the context window is worth a second look. Long-context marketing tends to be about how much you can put in, but on agentic work the output ceiling is usually what bites first, because a run that writes code, reads results and writes again accumulates output across turns.

The pricing structure also points to agent work. Output costs three times input, and cached input costs one eighth of fresh input. That ratio rewards a long, stable prompt reused across many short completions, which is the pattern of an agent working through a task against a fixed codebase or document set.

The numbers, and who produced them

Five benchmark results were published with the release. Read as a set, they are less flattering than the summary Alibaba led with.

The Terminal-Bench result is the one the company's framing rests on: 86.6 against 84.6 for two Anthropic models. Two points on a single agentic benchmark, self-run, is a thin basis for a claim about parity with frontier labs. The SWE-bench Pro result points the other way by a wider margin, 67.7 against 80.0, and that is Alibaba reporting its own gap.

The other three scores have no comparison figure attached. A number with nothing beside it tells a reader that the model was measured, not how it placed.

What Alibaba still has not said

The one missing specification determines what the model costs to serve: its activated parameter count. A 2.4-trillion-parameter mixture-of-experts model routes each token through a fraction of those parameters, and the size of that fraction is most of the inference economics. Alibaba has not disclosed it.

Some coverage has filled the gap with a figure. We are not going to, because the company that knows the number has chosen not to publish it, and an unattributed specification circulating in secondary reporting is not a specification.

The distinction is concrete. In a mixture-of-experts model the full parameter set has to be resident somewhere, which sets the memory bill, while only the activated fraction does arithmetic on each token, which sets the throughput bill. Two models with identical totals and different activation fractions have similar hardware footprints and very different serving costs. Without the fraction, a team can price Qwen3.8-Max on Alibaba's API and nowhere else, which is a reasonable commercial position for Alibaba to take and a real limit on what the weights release will be worth to anyone planning capacity.

The benchmark selection points the same way as the pricing. Terminal-Bench, OSWorld-Verified and SWE-bench Pro are all agentic or software-engineering harnesses, and QwenWork is an agent platform. A vendor's choice of benchmarks is a claim about what the model is for, and this set says the target is work that runs unattended, not conversation.

The availability notice carries a competitive fact as well as a technical one: Model Studio accepts OpenAI-compatible calls. For a team already running against that interface, evaluating this model is a change of base URL and key, not an integration project. Challenger models have leaned on that compatibility for two years now, and it is the reason a benchmark gap of a few points is worth arguing about at all — the cost of switching to find out is close to zero.

The open weights are the part that matters

A hosted model behind an API is a product. Published weights are a permanent fact, and they land in the middle of a policy argument that has no mechanism for them.

Four American labs met White House officials on 4 August over a voluntary framework that would give federal inspectors 30 days with a frontier model before release, and officials told the companies that open-weight models would not be included in that testing — a gap Anthropic pushed to close and did not. We reported that on 8 August: Washington Is Building a 30-Day Model Review.

The exclusion is not a loophole so much as a consequence of the mechanism chosen. Pre-release inspection needs a pre-release moment, and a weights release does not have one in any useful sense: once the file is distributed it is everywhere, permanently, in whatever modified form anyone wants. Qwen3.8-Max's weights are a live demonstration of that, arriving days after the framework was discussed.

What to do with a vendor benchmark

Three habits are worth keeping when a self-reported score arrives.

Read the whole published set, not the headline. Alibaba disclosed a twelve-point deficit on SWE-bench Pro in the same table as its two-point lead, and only one of those figures travelled.

Check whether a comparison figure exists at all. Three of these five scores have none, which makes them baselines for the next release rather than evidence about this one.

Then wait for a run somebody else controls. Terminal-Bench and SWE-bench Pro both have public harnesses, so independent numbers will follow, and the honest position until they do is that the vendor's claim is a claim.

What to watch

Whether the weights actually ship on the promised timetable, and under what licence. Whether independent runs reproduce the Terminal-Bench margin or erase it. And whether the activated parameter count is ever published, because until it is, nobody outside Alibaba can price serving this model on their own hardware.