Qwen 3.8

Canonical version: Qwen 3.8.

Qwen 3.8 is a 2.4 trillion parameter multimodal LLM from Alibaba's Qwen team, announced on 19 July 2026. Alibaba's claim is that it is "one of the most powerful models available today, compatible to leading frontier AI models, second only to Fable 5".

Two things matter more than that claim. The thing that shipped is qwen3.8-max-preview, not Qwen 3.8. And as of 1 August 2026, thirteen days after the announcement, there are still no weights, no model card, no license, and not one published benchmark.

What is actually available

• Qwen3.8-Max preview, live through Alibaba's Token Plan, plus the Qoder and QoderWork coding platforms
• Priced at 10% of standard pricing during preview. The standard per-token price has not been published
• 2.4T total parameters, sparse MoE, continuing the Qwen architecture line
• Active parameters per token: undisclosed. For a sparse MoE this is the number that determines both serving cost and real capability, and its absence is the most conspicuous gap in the announcement
• Multimodal: text, images, video, documents. The first Qwen multimodal model above 1T parameters
• Context window not stated
• Claimed to beat Qwen3.7-Max on coding, full-stack development, data analysis, and office workflows

What is not available

This list is the note. As of 1 August 2026:

• No open weights. Nothing from the Qwen organization on Hugging Face since June 2026. The only Hugging Face entries matching "Qwen3.8" come from unrelated third-party accounts publishing distills, which is odd given that no base weights exist to distill from. Treat those as unverified
• No benchmark table. Not the names, not the scores, not the prompts, not the harness, not the methodology
• No model card
• No license
• No standard API pricing
• No independent evaluation

Some secondary coverage states weights would land by 27 July. That is the Kimi K3 date and it appears to have been copied across. Alibaba committed to "soon" and to no date at all.

The timing is the point

The announcement came two days after Moonshot AI shipped Kimi K3 at 2.8T, with weights that actually appeared on Hugging Face. Read one way, Qwen 3.8 is a flag planted so the open-weight frontier does not belong to Moonshot uncontested for a fortnight.

What makes the promise genuinely interesting rather than merely defensive: Alibaba kept its flagship Max tier closed for two straight generations. Qwen3.6-Max and Qwen3.7-Max were API-only. Saying it will open the largest model it has ever trained is a reversal, not a continuation. That is worth watching precisely because it would cost Alibaba something real.

It is also, so far, a sentence rather than a repository.

Why it matters

The strategy shift is real, whatever this specific model turns out to be. Chinese labs built their reputation on value: small, fast, cheap, good enough. Qwen 3.8 and Kimi K3 are the opposite bet, big and slow and expensive to serve. Kimi K2 at 1T was the early signal a year ago; this is where it stops being one lab's choice and becomes the direction.

Commoditizing the model is a coherent way to attack the closed labs. If Alibaba Cloud is the best place to build on top of Qwen, the weights are a customer acquisition cost rather than the product. The Qoder and QoderWork bundling makes that reading explicit: the preview is gated behind Alibaba's own coding platforms, so the model is being used to sell the surface, not the other way round.

The counter-pressure worth naming. More open weights from one region can erode the incentive to train models anywhere else. Why spend the capex when someone gives it away? The rug-pull fear is overstated for weights you have already downloaded, and much less overstated for the training capability a country quietly stops maintaining. Open weights from several geopolitical regions is the outcome to want, and it is not the outcome currently trending.

Caveats

Heavier than usual, because there is almost nothing here to verify.

• Every performance claim is Alibaba's. No independent testing existed at the time of writing
• "Preview" is doing real work. Preview models get replaced, re-tuned, and re-benchmarked. Numbers gathered now may not describe what ships
• The relationship to the Qwen 3.6 and 3.7 lines is undisclosed. Whether 3.8 is a new architecture or a scale-up is unknown
• Big Chinese models in this generation are consistently reported as token-hungry. If 3.8 follows, effective cost will sit well above the headline price, and the headline price is not published either
• Secondary coverage is dominated by SEO content farms recycling the same announcement. Discount anything not traceable to Alibaba or to a hands-on test
• Revisit when weights land. Until then this note documents an announcement, not a model. The most informative future signal is simply whether the weights appear at all

References

• Alibaba Qwen announcement on X (2026-07-19) — https://twitter.com/Alibaba_Qwen/status/2078759124914098291
• Hacker News discussion (961 points) — https://news.ycombinator.com/item?id=48966120
• MarkTechPost, "Alibaba Previews Qwen3.8-Max, a 2.4 Trillion-Parameter Multimodal Model, Days After Moonshot's Kimi K3 Open-Weight Launch" (2026-07-19) — https://www.marktechpost.com/2026/07/19/alibaba-previews-qwen3-8-max-a-2-4-trillion-parameter-multimodal-model-days-after-moonshots-kimi-k3-open-weight-launch/
• Qwen Cloud token plan pricing — https://www.qwencloud.com/pricing/token-plan
• Qwen on Hugging Face (check here for weights) — https://huggingface.co/Qwen

Related

• Qwen
• Qwen3.6-35B-A3B
• Qwen3.6-27B
• Qwen Image 3.0
• Kimi K3
• Moonshot AI
• Deepseek
• Large Language Models (LLMs)
• AI Open Weight Models
• AI Frontier Model
• AI Foundation Models
• AI Mixture of Experts (MoE)
• Claude Fable 5
• GPT-5.6
• Gemini 3.6 Flash
• OpenRouter
• Artificial Analysis

About Sébastien

Ready to get to the next level?