Qwen3.8-Max is Alibaba's newest and largest Qwen model, announced on 2 August 2026 at 2.4 trillion total parameters. It is positioned around long-horizon agentic coding and office-style task execution — what Alibaba calls 'cowork' — rather than around conversational quality. At announcement it was available only as a closed preview, served as qwen3.8-max-preview through Alibaba's Token Plan, Qoder and QoderWork products. Alibaba stated that open weights would follow 'next week' but published no date, no licence, and no Hugging Face repository. The context window was not disclosed.

What benchmarks did Alibaba publish for Qwen3.8-Max?

None. At the time of the announcement Alibaba published no benchmark names, no scores, no evaluation harness, no prompts, and no methodology. This is unusual for a flagship release at this scale, and it is the single most important fact about the launch. Reporting circulated a claim that the model is second only to the current frontier leader, but there is no public benchmark table supporting that claim. Until numbers exist, the model is unevaluated rather than good or bad.

What did Alibaba claim instead of benchmarks?

Duration and autonomy. The announcement led with three anecdotes: more than ten days of self-evolving development taking a project from an empty folder to production without hand-holding, over 500 turns of chip design optimisation under system-level autonomous planning, and a 365-day e-commerce strategy loop. A GitHub project trace was offered as evidence for the first. These are claims about how long the model can run unsupervised, not about how well it answers any particular question.

Why is 'ten days of autonomous development' hard to verify?

Because there is no harness for it. A benchmark score is reproducible: you get the dataset, run the same prompts through the same scaffold, and compare. A ten-day agent trajectory is a single sample from a process with enormous variance, run inside an unpublished scaffold, on an unpublished task, with unpublished human intervention boundaries. You cannot rerun it, you cannot compare it against another model, and you cannot tell from the artefact how much of the result came from the model versus the surrounding orchestration. A GitHub trace shows what happened once. It does not show what happens on average, or what the same setup does with a different model.

Is Qwen3.8-Max worth using?

Unknowable today, which is the honest answer. It may well be excellent — Alibaba has shipped strong models repeatedly, and the agentic-coding focus is aimed at a real workload. But a 2.4-trillion-parameter model is also inference-infeasible for almost everyone outside a large datacentre, so even a future open-weights release would not make it locally runnable for most teams. The practical recommendation is to wait for the weights, the licence and a third-party evaluation before planning around it.

What is the more important part of the Qwen3.8 announcement?

Qwen3.8-27B. Alongside the 2.4T flagship, Alibaba said it would also open-weight a 27-billion-parameter model. That is the size class that fits on a single GPU, which means it is the one most developers can actually deploy, fine-tune and build products on. A 2.4T model is a statement about capability ceilings; a good 27B open model changes what a small team can ship. Historically the mid-size open Qwen releases have had far more downstream impact than the flagship.

How should teams evaluate the model when it does ship?

On your own workload, not on the vendor's. Assemble a set of tasks drawn from your actual codebase and business processes, fix the scaffold, and run the candidate model against your current model with everything else held constant. Measure completion rate, human intervention count, token cost per completed task, and failure mode — not just whether it eventually succeeded. For agentic use, run each task several times: variance between runs is often larger than the difference between two models, and a single successful trajectory tells you very little.

Qwen3.8-Max: 2.4 Trillion Parameters and No Benchmarks

Alibaba announced Qwen3.8-Max on 2 August. It is a 2.4-trillion-parameter model, the largest the Qwen team has shipped, and the announcement reached the second slot on Hacker News with 407 points and 165 comments.

It shipped with no benchmarks. No scores, no benchmark names, no harness, no prompts, no methodology. No licence, no repository, no disclosed context window. The preview is served through Alibaba’s own products — Token Plan, Qoder and QoderWork — as qwen3.8-max-preview, and open weights are promised for “next week” with no date attached.

That combination is the story. A flagship model at the largest parameter count any lab has publicly claimed, announced entirely on anecdote, and it worked: the post outranked almost everything else on the front page that day. What Alibaba replaced benchmarks with is worth looking at closely, because the replacement is spreading, and it makes the models harder to evaluate rather than easier.

Executive summary

Qwen3.8-Max: 2.4T total parameters, announced 2 August 2026, closed preview only, open weights promised without a date or licence.
Alibaba published zero benchmarks — no names, no scores, no methodology.
The headline claims are about duration: 10+ days of autonomous development from empty folder to production, 500+ turns of chip design optimisation, a 365-day e-commerce strategy loop.
A circulated claim that it ranks second only to the current frontier leader has no public benchmark table behind it. Treat as unverified.
Qwen3.8-27B is also going open-weights, and for most developers that is the consequential release.
The same week, DeepSeek published a full benchmark table for an open-weights model at 13B active parameters. The contrast in evidence standards is stark.

What was actually claimed

The announcement leads with three numbers, and none of them are scores.

The first is ten days. Alibaba says the model ran more than ten days of “self-evolving development,” taking a project from an empty folder to production without hand-holding, and points at a GitHub trace of the run. The second is 500 turns of chip design optimisation under what the post calls system-level autonomous planning with closed-loop adaptive learning. The third is a 365-day e-commerce strategy loop.

Read these carefully and they share a shape: each one is a claim about how long the model kept going, not about how well it did anything. Ten days of development is impressive if the output is production-grade and nobody intervened. It is meaningless if the agent spent nine of those days rewriting the same module, and there is nothing in the announcement that lets you tell the difference.

This is a deliberate positioning choice, not an oversight. Alibaba is selling into agentic coding and office automation, where the buyer’s question is “can I leave this running” rather than “what does it score on graduate physics questions.” Duration is the metric that maps to that question. The problem is that duration, as published, is not a metric at all.

Why a ten-day trajectory is not evidence

A benchmark score has one property that matters above all others: someone else can reproduce it. You take the dataset, you run the same prompts through the same scaffold, you get a number, and you compare it to the number the vendor published. When the numbers disagree, that disagreement is itself informative — it usually means the scaffold differs, and now you know something real about how the model behaves in your setup versus theirs.

A ten-day agent run has none of that. It is one sample from a process with very high variance. The task is unpublished, so you cannot attempt it. The scaffold is unpublished, so you cannot separate the model’s contribution from the orchestration around it — and in long-horizon agent work, the orchestration frequently matters more than the model. The intervention boundary is unpublished, so “without hand-holding” is a claim you have to take on faith rather than a measured quantity. And because it happened once, you have no idea whether the same setup succeeds nine times out of ten or one time out of ten.

The GitHub trace helps a little. It shows a real artefact was produced, which rules out the weakest form of the claim. It does not show what happens on the second attempt, and it does not show what a competing model does on the same task with the same harness, which is the only comparison anyone actually cares about.

There is a reason the field converged on shared benchmarks despite all their well-documented flaws. Benchmarks are bad at measuring capability and good at enabling comparison. Anecdotes are the reverse: vivid about capability, useless for comparison. Replacing the former with the latter is not a step forward in evaluation rigour, whatever it does for the marketing.

The contrast, in the same week

The strongest argument against grading Qwen3.8-Max generously is that another lab demonstrated the alternative days earlier.

DeepSeek released V4 Flash 0731 on 31 July: 284 billion total parameters with 13 billion active, a one-million-token context window, MIT licensed, weights on Hugging Face. It shipped with a published table.

Benchmark	Previous V4 Flash	V4 Flash 0731
Artificial Analysis Intelligence Index	40	50
Terminal-Bench 2.1	61.8	82.7
DeepSWE	7.3	54.4
GDPval-AA v2 (Elo)	1189	1559

Those numbers are checkable. They sit on a third-party index, they name the benchmarks, and anyone with the weights and the harness can attempt to reproduce them. The Terminal-Bench jump in particular — more than twenty points on an agentic terminal-use benchmark — is the kind of claim that invites scrutiny precisely because it is specific enough to be wrong.

Both labs are making a claim about agentic capability. One made it in a form that can be checked and, if wrong, corrected. The other did not. That difference is not a technicality; it is the whole basis on which a team decides what to deploy.

The 27B is the real release

Buried under the 2.4T headline is the announcement that Qwen3.8-27B is also going open-weights.

For most people reading this, that is the entire story. A 2.4-trillion-parameter model is a statement about what is possible at the ceiling. Even fully open-weighted, it is not something a startup runs — the serving footprint puts it firmly in datacentre territory, and the practical access path remains someone else’s API. A 27-billion-parameter open model is different in kind. It fits on a single GPU, it can be fine-tuned on a modest budget, it can be embedded in a product without a per-token bill, and it can be audited.

The historical pattern supports this. Across open-model families, the mid-size releases consistently generate more downstream work — more fine-tunes, more quantisations, more shipped products — than the flagships that get the announcement. The flagship sets the marketing anchor. The mid-size model is what people build on.

If you are deciding where to spend attention when the Qwen3.8 weights land, spend it there.

What this means for how you evaluate models

The broader shift is real and it is not confined to Alibaba. Long-horizon autonomy is genuinely becoming the workload that matters, and the standard static benchmarks measure it badly. Andrej Karpathy made the same point from the other direction this week, arguing that single-shot prompt tests have outlived their usefulness and demonstrating the alternative with a multi-hour, million-token generation task. He is right that the old tests are saturating.

But “the old measurements are inadequate” does not license “publish no measurements.” It creates an obligation to build better ones, and the gap between those two responses is where a lot of 2026 model marketing now lives.

For teams, the practical consequence is that vendor claims are getting less useful, not more, and the burden shifts onto you. A workable approach:

Build a task set from your own work. Twenty to fifty real tasks from your codebase and processes, with a clear definition of done. This is the only benchmark that predicts anything about your deployment.
Hold the scaffold fixed. Change one variable at a time. If you swap model and framework together, you learn nothing about either.
Run each task multiple times. Agentic variance between runs routinely exceeds the difference between two models. A single trajectory — yours or a vendor’s — is an anecdote.
Measure cost and intervention, not just success. A model that completes the task for four times the tokens and two human rescues has not completed the task.
Record failure modes. How a model fails is more actionable than whether it failed, and it is the thing no benchmark score will ever tell you.

The verdict

Qwen3.8-Max might be very good. Alibaba has shipped competitive models repeatedly, the agentic-coding focus targets a real and valuable workload, and the willingness to open-weight at all remains a meaningful contribution to the ecosystem.

None of that is a reason to accept the announcement as evidence. As published, it is a positioning document: a large number, three unreproducible anecdotes, and a promise of weights. The correct response is to note it, wait for the licence and the numbers, and in the meantime pay attention to the 27B — and to the models that showed their work.