Alibaba Answered Kimi K3 in Three Days: Qwen3.8 Is 2.4 Trillion Parameters, Going Open-Weight, and Documented in One Tweet

TL;DR

On Sunday, Alibaba's Qwen team announced Qwen3.8: a 2.4-trillion-parameter flagship that is "launching and going open-weight soon," with a preview build, Qwen3.8-Max-Preview, already live behind its Token Plan subscription and inside its Qoder coding tools. Alibaba's own ranking puts it "second only to Fable 5." What Alibaba has not published: benchmarks, a license, a release date, or a model card. Three days after Moonshot AI claimed the biggest-open-model crown with the 2.8T Kimi K3, China's open-weight race has a second frontier-scale checkpoint promised in the same month, and the entire spec sheet for this one currently fits in a tweet, with room left over for the globe emoji.

One Tweet, 2.4 Trillion Parameters

The announcement is short enough to quote nearly in full. Qwen3.8 is "launching and going open-weight soon," carries "a massive 2.4T parameters," is "continuously evolving," and sits, in Alibaba's telling, "second only to Fable 5" among models available today. That last phrase is doing a lot of load-bearing work: it is a claim about the frontier, delivered with zero published evals attached.

Here is what is concretely true right now. Qwen3.8-Max-Preview is a real endpoint you can use today. It is live on Token Plan, Alibaba's model subscription tier (pitched at roughly 40% off pay-as-you-go rates), and inside Qoder and QoderWork, its agentic coding products. Access is subscriber-gated: there is no public per-token price sheet for the new model and no open API listing yet.

Note the pairing of "preview" with "continuously evolving." Alibaba is saying out loud that the checkpoint you poke this week is not final. Any vibes-based eval you run against it today is provisional by design, and the weights that eventually land may not be the model you tested.

The Week the Open-Weight Ceiling Moved

Zoom out five days and the pattern is hard to miss. On July 15, Thinking Machines shipped Inkling, a 975B Apache 2.0 release with weights you can download right now. On July 16, Moonshot announced Kimi K3 at 2.8 trillion parameters, calling it the largest open-source model to date, weights promised for later this month. In mid-July, DeepSeek graduated V4 to official release; the 1.6T V4-Pro weights have sat on Hugging Face under MIT since the April preview. And now Qwen3.8 at 2.4T. A year ago, one trillion-parameter open checkpoint was an event. This week produced a queue of them.

The escalation reads as directly competitive, and the marketing copy gives it away. K3's launch line was that it trails only Fable 5 and GPT-5.6 Sol; Qwen's, three days later, is that it trails only Fable 5. "Second only to Fable 5" is becoming the "objects in mirror are closer than they appear" sticker of Chinese model launches: Anthropic is now the unit of measurement, and each new release rounds itself one slot closer.

Alibaba has extra reasons to move fast. Qwen was just cleared as the model behind Apple Intelligence in China, so this is not a research lab flexing; it is the default assistant for a nine-figure install base defending its home turf.

What "Open" Means at 2.4 Trillion Parameters

Napkin math first. A 2.4T-parameter checkpoint is roughly 4.8 terabytes of weights at FP16, about 2.4TB at 8-bit, and still on the order of 1.2TB quantized down to 4 bits, before you allocate a single byte of KV cache. No consumer GPU, and no reasonable homelab, is in this conversation.

An open checkpoint at this scale is like being gifted a free 747: legally yours, genuinely valuable, and entirely dependent on someone else's hangar, fuel, and crew. In practice, frontier-scale open weights get consumed three ways: served by inference providers, fine-tuned or distilled by companies with cluster budgets, and audited by researchers. The download-and-run-local crowd is not the audience, and the Hacker News thread knows it: the recurring request is for 27B-to-122B variants, the sizes that fit on hardware people actually own. Alibaba has said nothing about smaller siblings.

That does not make the pledge meaningless for builders. Every serious hosting provider can serve the same frontier-class checkpoint once it drops, so the competition moves to price, latency, and reliability rather than model access. It is no accident that the cheapest frontier-class APIs right now sit on top of open or opening checkpoints: DeepSeek's MIT-licensed V4-Pro serves at under a dollar per million output tokens off-peak. A 2.4T Qwen in that pool applies the same pressure from a second direction.

The Missing Paperwork

Before treating Qwen3.8 as more than a promissory note, watch for four things:

• The license. The Qwen3 generation of open releases shipped under Apache 2.0. If 3.8 follows suit, it becomes one of the largest permissively licensed models ever published. If a custom license shows up instead, that is its own story.
• The date. "Soon" is not a date. Moonshot at least attached a window to K3's weights; Alibaba attached an emoji.
• The card. Architecture, active parameters, context window, training data: all unpublished. At this scale a sparse mixture-of-experts design is the overwhelming norm, but that is inference on our part, not disclosure on theirs.
• The delta. "Continuously evolving" means the preview and the eventual open weights may differ. Whether the community gets the model in the demo or a smaller, older cousin is the difference between a landmark release and a press cycle.

Key Takeaways

• Alibaba announced Qwen3.8 on Sunday, July 19: 2.4 trillion parameters, "going open-weight soon," with Qwen3.8-Max-Preview already live on Token Plan, Qoder, and QoderWork.
• The launch shipped no benchmarks, no license, no release date, and no model card; the "second only to Fable 5" ranking is a vendor claim with no published evals behind it.
• It is the second frontier-scale open-weight commitment from China in three days, after Kimi K3's 2.8T, and the fourth big open-weight move in a week counting DeepSeek V4 and Inkling.
• At 2.4T parameters the checkpoint is on the order of 1.2TB even at 4-bit, so the real consumers are hosting providers, fine-tuning shops, and researchers, not local rigs; the community is already asking for 27B-122B variants.
• The license and the actual drop date are the things to watch; open frontier checkpoints shift competition to hosting price and latency, where DeepSeek is already fighting.

Sources: Alibaba Qwen announcement on X, Yahoo Finance: Alibaba's Qwen unveils preview of flagship AI model, QwenCloud Token Plan, Hacker News discussion