Alibaba has unveiled Qwen 3.8, a 2.4 trillion parameter model its team frames as trailing only Fable 5. It is the first multimodal model in the Qwen family to clear the one trillion mark, and it landed twenty-four hours after Kimi K3.
Key Takeaways
- Qwen 3.8 carries 2.4 trillion parameters and handles images, videos and documents.
- No benchmark results have been published so far.
- The preview runs at 10 percent of the standard price, with open weights promised but undated.
A 2.4 trillion parameter model with no numbers attached
Alibaba framed Qwen 3.8 as a model that now matches the frontier, placing it directly behind Fable 5 in the current pecking order. That is a loud claim. It arrived with no results table.
The gap is the first thing worth flagging. A lab that says it sits one rung below the best model on the market usually ships the numbers that prove it, and ships them the same day. The Qwen team announced the ranking before the evidence.
The 2.4 trillion figure still says something. Qwen 3.8 is the first multimodal model from the house to break the trillion barrier, pulling images, videos and documents through the same pipeline.
The timing explains itself. Moonshot had just shipped Kimi K3 and its 2.8 trillion parameters, a release we covered last week when the largest open model of the moment went after the top American systems. Alibaba answered inside a day.
That pace tells you where the Chinese market is. Two labs pushed out their heaviest models a day apart, precisely while the Shanghai conference held the international spotlight.
The reaction spread well past developer circles. The Kimi K3 launch knocked the Nasdaq down about 1 percent on semiconductor selling, a sign that markets now read these announcements as industrial news rather than technical curiosities.
For an engineering team the missing numbers create an immediate practical problem. There is no way to estimate cost per task, latency or failure rate without running your own evaluation, which pushes onto the customer the assessment work the lab chose not to publish.
A preview at a tenth of the price, sold before the weights land
Access to Qwen 3.8 runs through a preview billed at 10 percent of the standard price, served through Alibaba’s Token Plan alongside Qoder and QoderWork. The discount is aggressive and deliberate.
For a product team the maths is blunt. A multimodal model this size at a tenth of the usual rate makes testable what was not, especially on video workloads that stayed out of budget with most vendors.
The catch sits in the same sentence. With no benchmarks, nobody yet knows how the model behaves on long reasoning or on code, and a subsidised preview says nothing about the price once the promotion ends.
Open weights are promised, with no date attached. Until they ship, Qwen 3.8 remains a proprietary model reachable through an API, and the openness pledge works mostly as a signal aimed at the community.
Alibaba has form here. The group had already cut its own staff off from Claude Code to push its in-house tooling instead, a call that traced the same software sovereignty logic.
The wider shift reads better on video, and the move by Western buyers toward Chinese models on pure cost grounds is exactly what this short piece on the Coinbase switch walks through.
That makes the weight release schedule the variable that matters. While it stays vague, a team building on Qwen 3.8 is accepting dependence on an API whose promotional rate can move without notice, with an exit cost that climbs the deeper the integration goes.
Multimodality changes the arithmetic too. Video eats far more compute than text, and a discount of this size on that class of workload can revive entire projects that were sitting idle for want of an inference budget.
More articles on Horizon
- Gemini 3.5 Pro Delayed Again as Google Rebuilds It
- Kimi K3 Closes the Gap With Top US AI Models
- Anthropic Cuts Fable 5 Limits Starting July 20
What answering Kimi K3 in a day means for Western labs
Two models above two trillion parameters in two days, both Chinese, both announced with an openness pledge. Release cadence has become a weapon of its own.
For Western labs the pressure has moved off raw capability. It sits on price, and a preview at a tenth of the standard rate plants a number that buyers will carry into their next renewal conversation.
There is precedent. An open Chinese model crossed that line when MiniMax M3 edged past GPT-5.5 in open access, and that episode showed how fast an aggressive price-to-quality ratio moves real usage.
One question stays open today. With no published figures and no downloadable weights, the claimed slot behind Fable 5 is a commercial statement rather than a verifiable result.
There is a fresh precedent for that caution. Kimi K3 shipped with scores attached, and those same numbers showed a clear lead on frontend code sitting next to a severe drop on hard mathematics. A model can top one board and remain useless on the task you actually care about.
On the supply side the sharpest consequence lands on pricing structure. When two very large models arrive in subsidised preview in the same week, the market reference price falls before anyone has settled the quality question.
Usage will settle it over the coming weeks. Teams wiring Qwen 3.8 into real workloads will produce the comparisons the lab withheld, and that is when the claim either holds or breaks.
Follow the story on Horizon.
