Qwen3.8-Max Ships: 2.4T MoE, Open Weights Still Pending

Qwen3.8-Max is out — for real this time. On August 3, 2026, Alibaba moved its biggest model ever from teaser to general availability: a 2.4-trillion-parameter sparse Mixture-of-Experts that activates just 95 billion parameters per token, takes text, image, and video input, and carries a headline 1M-token context window.

The release answers most of what the July preview left open. There is now a full vendor benchmark table — roughly 40 rows with 20 numbered methodology footnotes — an active-parameter disclosure, a live hosted API with drop-in snippets for popular coding agents, and a dated commitment to something Qwen has never done before: open-sourcing the weights of a Max-class flagship, promised for the following week alongside a new 27B model.

This guide separates what verifiably shipped on day one from what is still a promise, reads the vendor benchmark table with its own footnotes in hand, and covers the pricing that is circulating but not yet on Alibaba’s own price page. Everything below is sourced from the Qwen Team’s release post and same-day coverage, and every score is vendor-stated unless noted.

The Qwen Team’s release post opens without hedging: this is the official release of the most capable model in the Qwen family, not a staged preview. The model is a sparse Mixture-of-Experts design with hybrid attention, built on the architectural foundation of Qwen 3.5, and the post states the activation figure plainly — despite 2.4 trillion total parameters, only 95 billion are active per token. Input is multimodal (text, images, and video), output is text, and the headline context window is one million tokens.

That inverts the July situation. When Alibaba teased this model at WAIC Shanghai, there were no benchmarks, no pricing, and no open-weight date — our coverage of the July 19 preview treated the 2.4T claim as exactly that, a claim. The full release fills in two of those three gaps with published detail and converts the third into a dated commitment. The August 3 headlines treated it as a market event too: Bloomberg’s headline framed the release around benchmark claims rivaling Anthropic, and CNBC’s led with an Alibaba share rally.

The one same-day write-up we could read in full got the “what has been published” question outright wrong — as we detail in Section 04. The table below is our own audit of the release post: every claim sorted into shipped, promised, or undisclosed, based on what is actually present on (or absent from) the page.

The pattern this table exposes is the story: Alibaba shipped the proof points that make headlines — benchmarks, specs, API — on day one, and deferred the two items that determine whether enterprises can actually build on this model long-term: the weights and the license. Both of those now have a clock running on them, which is exactly why they deserve tracking rather than assumption.

Every figure in this section comes from Alibaba’s own benchmark table on the release post — self-reported scores, scored against Claude Opus 4.8, Claude Fable 5, GPT-5.6 Sol, Gemini 3.1-Pro, and its own predecessor, Qwen3.7-Max. Treat them as the vendor’s case, not a neutral evaluation. Read that way, the table is more interesting than a clean sweep would be: Alibaba published rows where its model loses, and the wins cluster in a specific place.

Alibaba published individual rows with no roll-up, so we computed the margins ourselves. In the table below, every margin is our own arithmetic on the vendor’s published rows: the Qwen3.8-Max score minus the strongest rival score on the same row.

The shape is consistent. Qwen3.8-Max’s wins sit in computer use, research-workflow reproduction, and instruction following — the agentic-operations cluster. Its deficits sit in repo-scale software engineering, where the vendor table itself shows double-digit gaps to Claude Fable 5 on SWE-bench Pro and FrontierSWE. Elsewhere in the table Alibaba reports a GPQA Diamond of 92.6, and the release cites rankings on Alibaba’s own internal arenas (5th in its Text Arena, 2nd in Vision, 4th in Frontend Code) — vendor-run leaderboards, not an independent community ranking.

The generational jump is real even on the rows it loses. On DeepSWE 1.1, Qwen3.8-Max scores 56.6 against its predecessor Qwen3.7-Max’s 21.6 — a 35-point improvement in one generation on the vendor’s own numbers, even though it still trails GPT-5.6 Sol’s 73.0 on the same row. A model family that closes gaps at that rate is worth re-evaluating every cycle regardless of where it stands today.

Day-one coverage of this release is a case study in why primary sources matter. MarkTechPost’s same-day write-up claimed: “No benchmark table, license, or activated-parameter count has been published.” Two of those three claims do not survive contact with the release post itself, which carries a full benchmark table under an explicit heading and states the 95B active-parameter figure twice. Only the license claim holds — no license text appears anywhere on the page.

Neither caveat is a scandal — disclosing your harness in 20 numbered footnotes is better practice than most launches manage. But together they define how to use this table: as Alibaba’s best case under favorable budgets, against a baseline of its own choosing on the multimodal rows. The gap between a 5-hour-timeout benchmark score and what your agent does under a production time budget is precisely the gap your own evaluation has to measure.

The circulating price for Qwen3.8-Max is $2.00 per million input tokens and $6.00 per million output tokens — a standard hosted-API rate, with reported cache tiers of $0.25 per million for implicit cached input, $2.50 for explicit cache creation, and $0.17 for explicit cache reads. Multiple independent write-ups converge on those numbers to the cent, which suggests a real vendor source behind them. But we could not find any qwen3.8-listed rate on Alibaba’s own public Model Studio pricing page at the time of writing — its Qwen section listed model IDs only through the qwen3.7 family. Treat the figures as reported, not as a confirmed vendor list price, and note the cache tiers are the least-corroborated of the set.

The operational trap is the default. With reasoning_effort shipping at xhigh and thinking preserved across turns, the out-of-box configuration is tuned for benchmark-grade output, not for cost. Before comparing this model’s economics against your incumbent, decide the effort level per workload class — and if you are already inside Alibaba’s ecosystem, reconcile these per-token rates against Alibaba’s token-plan pricing structure, which changes the effective math for prepaid quota.

The most distinctive part of the release post is not the benchmark table — it is three long-horizon autonomy case studies. All three are Alibaba’s own showcases and should be labeled that way, but two of them left public artifacts that anyone can inspect, which is more receipts than launch-day demos usually offer.

The research-reproduction study comes with a round-by-round table on the release post, which we reproduce here with attribution — these are Alibaba’s numbers, not ours:

A fourth showcase is fully in-house: on Alibaba’s E-Commerce Bench — a 365-day simulated store operation starting from ¥100,000 in capital, spanning 12 store types, roughly 600 suppliers, 7,000 products, and 152 embedded fraudulent suppliers — Qwen3.8-Max finished at ¥416,252, a 4.16x return that Alibaba reports as 38% ahead of second-place GLM 5.2 and a 152% improvement over Qwen3.7-Max. A vendor benchmarking rivals on its own simulation deserves the heaviest discount of anything in the release. The more durable signal is the format itself: publishing inspectable, long-horizon artifacts — a public repo, a contest leaderboard — is becoming how labs argue for agentic capability, and it is a better argument than any static table row.

The release’s biggest strategic move ships nothing at all on day one. Qwen has never open-sourced a Max-class flagship — its 2026 pattern has been closed flagships, open everything else — and the release post commits to breaking that pattern within a week, for two models at once: Qwen3.8-Max itself and a new, otherwise-undescribed Qwen3.8-27B.

The missing piece is the license. The release post contains no license text at all, and the license determines everything about what “open weights” is worth: commercial-use rights, redistribution, fine-tune ownership. Qwen’s 3.5 and 3.6 open releases shipped under Apache 2.0, which sets a reasonable expectation — but that is precedent, not a fact about 3.8, and a promise with no license attached is not yet something to build a procurement decision on. The contrast with Moonshot’s K2.7-Code release is instructive: that model shipped weights and license on launch day, while Qwen is running a closed-then-open sequence with a one-week gap. We cover exactly what to verify when the weights actually land — checksums, license text, config parity with the hosted model — in our pre-download checklist for the Qwen3.8 weights.

Alibaba made trialing this model unusually cheap. The release post includes official drop-in configurations for Claude Code, Codex, Qoder CLI, Qwen Code, and OpenClaw — pointing Claude Code at Qwen3.8-Max is two environment variables (ANTHROPIC_BASE_URL set to Alibaba’s Anthropic-compatible endpoint and ANTHROPIC_MODEL=qwen3.8-max), and the new Qwen-MM-Plugins library extends existing agent harnesses with image, video, and visual-tool capabilities. The evaluation cost is an afternoon, not a migration.

The trend this release confirms is that the open-weight frontier is now a staged-release game. Chinese labs in particular have spent 2026 converging on the same play: launch closed with a benchmark case, monetize the API window, then open the weights once the news cycle has done its work. Qwen running that sequence on its Max-class flagship — the tier it always kept closed — is the strongest version of the pattern yet, and it puts pressure on every lab whose open-weight story stops below flagship scale.

Looking forward, the week of August 10 is the real test. If both checkpoints land under a permissive license, a Max-class frontier model becomes self-hostable for the first time in Qwen’s history, and the 27B gives mid-size teams a fine-tunable sibling from the same generation. If the weights slip or the license arrives with meaningful restrictions, this release will be remembered as a benchmark story instead. For teams deciding how — or whether — to fold a model like this into production pipelines, that is a routing-and-governance question before it is a benchmark question, and it is exactly the kind of comparative evaluation our AI transformation engagements are built around.