Qwen3.8's 2.4T Open-Weights Promise: Now With a Date, Still Without a License

Last updated: August 2026

On July 19, 2026, Alibaba's Qwen team announced Qwen3.8: a 2.4 trillion parameter model it describes as "one of the most powerful models available today," second only to Anthropic's Claude Fable 5, with an open-weight release promised "soon." The announcement landed exactly three days after Moonshot AI launched Kimi K3 at 2.8 trillion parameters with a committed weight-release date of July 27. Two trillion-scale "open" flagships from two Chinese labs inside one week is not a coincidence, and the relationship between the two companies makes the timing stranger still.

This site covered how Kimi K3 compares to the closed flagships when it launched. Qwen3.8 cannot get the same treatment: there is still nothing independent to measure. What can be examined right now is the gap between what Alibaba announced and what Alibaba published, the track record behind the open-weight promise, and what any of this changes for someone running AI on hardware they own.

Update, August 2, 2026: Alibaba has committed the weights to next week, published per-token pricing, and announced a smaller Qwen3.8-27B going open-weights alongside the flagship. The checklist at the bottom of this page is scored accordingly, and our new companion piece runs the hardware reality check on both models.

What Alibaba Announced, and What It Did Not

Strip the announcement to its verifiable parts and three things remain. First, Qwen3.8-Max-Preview is a real, purchasable product, live on Alibaba's Token Plan, Qoder, and QoderWork as of July 19. Second, Alibaba states the model has 2.4 trillion total parameters. Third, Alibaba says the full Qwen3.8 will be released with open weights.

Now the list of what was not published, which is longer and matters more. There is no active-parameter count and no mixture-of-experts configuration. There is no license text. The release date graduated from open-ended to a committed week in the August 2 follow-up. There is no model card on Qwen's Hugging Face organization as of August 2, 2026, which is the clearest signal that nothing has actually shipped. There is no benchmark table of any kind, which means the "second only to Fable 5" claim is a marketing sentence, not a measurement. Every one of those artifacts exists for Kimi K3 in some form; none of them exists for Qwen3.8.

One more line in the announcement deserves attention: Alibaba describes the preview as "continuously evolving." That is vendor language for a model that can change underneath you without notice. It is a normal practice for hosted previews, and it is also precisely the failure mode that owning weights eliminates. A model whose behavior shifts on the vendor's schedule is the rental-equipment problem this site was built around, applied to intelligence instead of modems.

The Open-Weight Promise Has Two Recent Counterexamples

Qwen earned its reputation with developers through genuinely open releases. The Qwen3, Qwen3.5, and Qwen3.6 open lines ship under Apache 2.0, one of the most permissive licenses in the industry, and models like Qwen3.6-35B-A3B are staples of the self-hosting community. That history is real, and it is why the open-weight claim is plausible rather than absurd.

But the flagship tier tells a different story. Qwen3-Max-Preview, Alibaba's first trillion-parameter model, launched in September 2025 as a hosted product. Qwen3.6-Max-Preview followed in April 2026 and stayed closed, available only through Alibaba's cloud platforms. Qwen3.7-Max arrived in May 2026 the same way: API-only, no weights published. To be fair to Alibaba, neither of those releases came with an open-weight promise that was then withdrawn; the company simply shipped its biggest models closed while its smaller lines stayed open. That is exactly why the Qwen3.8 announcement is notable. If the weights actually ship, Alibaba is reversing its own flagship strategy, not continuing it. Until the files appear on Hugging Face with a license attached, the honest status of Qwen3.8 is: a closed hosted model with a dated commitment attached, and a commitment is still not a download.

The In-House Race: Alibaba vs. the Lab Alibaba Backs

The three-day gap between the Kimi K3 launch and the Qwen3.8 announcement reads like direct competition, and it is. What makes it unusual is that Alibaba disclosed in its fiscal 2024 annual report, filed with the Hong Kong stock exchange, that it had invested approximately 800 million dollars for an approximately 36 percent equity stake in Moonshot AI, as reported by the South China Morning Post. Later funding rounds have diluted that position, and Moonshot was valued around 4.8 billion dollars in a round reported by CNBC in January 2026, but Alibaba remains the company's leading corporate backer.

So the week's sequence is: a lab announces the largest open-weight model ever, with a hard date and disclosed architecture, and its biggest corporate shareholder responds within seventy-two hours by teasing a slightly smaller model with no date and no architecture, timed to land before the first lab's weights do. Whatever the boardroom logic, the effect for readers is useful: two heavily resourced labs are now competing on who can put frontier-scale weights into public hands, and public-weight competition is the kind that compounds in your favor. The pressure that matters here is not between countries; it is between release strategies. Open publication forces verification, invites hosting competition, and survives its vendor. Closed hosting does none of those things, whichever company is doing it.

The Only Honest Comparison Table Anyone Can Publish Today

A benchmark comparison between Qwen3.8 and anything else is impossible right now, because Qwen3.8 has published no benchmarks. What can be compared, cell by cell, is what each model has actually put on the record. The empty cells below are the story.

Status as of August 2, 2026. Kimi K3's weights, model card, license, and technical report shipped July 27; broad third-party evaluation of the released model is still limited. Qwen3.8 figures are Alibaba's announcement claims.

Read across the Qwen3.8 row and the argument makes itself. On this table, the model Alibaba says is second only to Fable 5 currently has less on the public record than the closed models it is supposed to be an open alternative to. Fable 5 does not publish its parameter count either, but it also does not ask for credit as an open release. The moment you claim "open," the empty cells become your problem.

The Memory Math Nobody at Alibaba Mentioned

Set the disclosure gaps aside and assume the best case: weights ship, permissive license, sane architecture. The arithmetic still puts Qwen3.8 nowhere near a home machine. At 4-bit quantization, 2.4 trillion parameters is roughly 1.2 terabytes for the weights alone, before context cache and runtime overhead. The largest consumer-reachable memory pool today, a 512GB Mac Studio, is not half of it.

Whether the model is sparse or dense changes serving cost dramatically, which is why the undisclosed active-parameter count is the single most consequential missing number. DeepSeek V4-Pro shows what disclosure looks like: 1.6 trillion total, roughly 49 billion active per token, about 3 percent of the network, which is the entire reason it serves at a rational cost. A sparse 2.4T and a dense 2.4T are different machines to run, and today nobody outside Alibaba knows which one Qwen3.8 is. We worked through this exact wall for Moonshot's model in our Kimi K3 hardware reality check; every number there applies here with the dial turned down about 14 percent, and the conclusion does not change.

What This Race Delivers to You Anyway

None of this means the announcement is irrelevant to someone running AI at home. Frontier-scale open releases pay out to consumers through three channels, none of which require you to host the model.

The first is price. When frontier weights go public, any datacenter can serve them, and hosted inference becomes a commodity market instead of a vendor's margin. Kimi's launch pricing already came in far below the closed flagships, and every additional open release at this tier adds pressure; Alibaba's August 2 per-token pricing for Qwen3.8-Max, $2.00 per million input and $6.00 per million output, continues the trend. The second is permanence. A model whose weights you or anyone else can store cannot be deprecated out from under the people using it; a hosted preview described as "continuously evolving" is the opposite of that guarantee. The third, and the one that reaches your hardware, is distillation. Trillion-scale open models become the teachers for the 8B-to-70B models that actually fit on machines people own. The strong small models in our guide to the best local AI models by VRAM tier exist in large part because bigger open models existed first.

A fourth point is easy to lose in a parameter-count arms race: the models you can run today are already good. A 16GB machine runs genuinely capable models right now, and a dedicated always-on box is affordable; our mini PC guide for local AI covers the hardware tiers and the network isolation that should come with them. Waiting for a 1.2-terabyte download to become practical costs you nothing while everything below it keeps improving. Meanwhile, using Qwen3.8-Max-Preview or any hosted preview means your prompts route through the provider's infrastructure under the provider's terms, the tradeoff we documented clause-by-clause for Kimi K3's own policies.

What to Watch: Four Artifacts That Turn Claims Into Facts

This story resolves on evidence, not announcements, and the evidence has a short checklist. Watch for the license text, because "open-weight" means nothing specific until the terms are published, and recent open releases have carried commercial-use clauses worth reading. Watch for the active-parameter disclosure and architecture details, because that number decides whether the model is serveable by anyone besides Alibaba. Watch for the model card and weight files on Hugging Face, because that is the moment a promise becomes a download. And watch for independent benchmarks, because until third parties test the released model, the ranking claim remains an advertisement.

The clock resolved in Moonshot's favor. Kimi K3's weights, license, and technical report shipped July 27, making it the largest open-weight model ever released, while Qwen3.8 stayed a teaser for another week. Alibaba answered on August 2: a committed window of next week, published per-token pricing, and Qwen3.8-27B going open-weights alongside the flagship, the one model here sized for hardware you own.

Scoring our own checklist: the date and the pricing have landed. The license, the active-parameter disclosure, the model card, and the independent benchmarks have not; the new August 2 benchmark chart is entirely vendor-run. The weights remain the scoreboard. For the hardware math on both models, from the 1.2-terabyte flagship to the 27B that should fit a 24GB GPU, our Qwen 3.8 hardware reality check runs the numbers. This page gets its final scoring when the files exist.

Frequently Asked Questions

Is Qwen3.8 open source?

No, not today. Qwen3.8 is available only as a hosted preview on Alibaba's platforms. Alibaba has committed an open-weight release for next week, but no weights or license have been published as of August 2, 2026. Note that even when weights ship, "open weight" means downloadable parameters, not full open source with training data and code.

When will Qwen3.8's weights be released?

Next week, per Alibaba's August 2, 2026 announcement, which also committed Qwen3.8-27B to open weights. No exact day, license, or repository has been published; the release is real the moment the model card and weight files appear on Hugging Face, and not before.

Can you run Qwen3.8 locally?

Not the 2.4T flagship: at 4-bit it needs roughly 1.2 terabytes for the weights alone, far beyond any consumer machine. The realistic local target is the newly announced Qwen3.8-27B, which should fit a 24GB GPU or 32GB Mac at 4-bit if it follows the Qwen 27B line's sizing. Our Qwen 3.8 hardware reality check runs the full numbers for both models.

Is Qwen3.8 better than Kimi K3?

Nobody outside Alibaba can say, because Qwen3.8 has published no benchmarks. Kimi K3 has at least partial third-party results and shipped its weights, license, and technical report on July 27, making it the far more verifiable release. Treat any Qwen3.8 ranking claim, including Alibaba's own "second only to Fable 5," as unverified marketing until independent testing exists.

What is Qwen3.8-Max-Preview?

It is the hosted preview variant of Qwen3.8, live on Alibaba's Token Plan, Qoder, and QoderWork since July 19, 2026. Alibaba describes it as "continuously evolving," meaning the model can change during the preview window and may be swapped or retired when the production version ships.

Is Qwen3.8 the biggest AI model ever?

No. Kimi K3, announced three days earlier, is larger at 2.8 trillion parameters, and closed frontier models do not disclose their sizes at all. K3's weights have been downloadable since July 27, 2026, which makes it the largest open-weight model ever released. If Qwen3.8 ships next week as committed, it becomes the second-largest open release, well ahead of DeepSeek V4-Pro at 1.6 trillion parameters.

What can I actually run at home right now?

A great deal. Consumer-class open models in the 8GB-to-128GB range are genuinely capable in 2026 and keep improving on hardware you already own. Our best local AI models by VRAM tier guide matches a model to your memory; our mini PC guide covers a dedicated, network-isolated box for the job.