Alibaba announced the Qwen 3.8 open weight model on July 19, 2026 at the World AI Conference in Shanghai — a 2.4-trillion-parameter model the company says will receive open weights “soon” — and as of July 22, 2026 the weights have not shipped. Only Qwen3.8-Max-Preview is available, running on Alibaba’s Token Plan, Qoder, and QoderWork platforms.
The Qwen team published the announcement through its official @Alibaba_Qwen account. Qwen3.8 uses a sparse mixture-of-experts architecture and totals 2.4 trillion parameters. Alibaba states the model ranks “second only to Fable 5” among frontier systems, based on internal evaluation. Alibaba has published no benchmark table, no license, no model card, and no open-weight release date.
What Alibaba Confirmed About the Qwen 3.8 Open Weight Model
Alibaba confirmed 5 concrete attributes of Qwen3.8: 2.4 trillion total parameters, a sparse mixture-of-experts architecture, multimodal input, preview availability on 3 Alibaba platforms, and a promised open-weight release with no date attached. Everything else about the model remains undisclosed.
“Qwen3.8 is launching and going open-weight soon! With a massive 2.4T parameters, this model is continuously evolving. We believe it’s one of the most powerful model available today, compatible to leading frontier AI models, second only to Fable 5.” — @Alibaba_Qwen, July 19, 2026
According to Shuai Bai, a developer on the Qwen team, Qwen3.8 is “the team’s first multimodal model with more than 1 trillion parameters.” The preview accepts both OpenAI and Anthropic API protocols, so existing application code calls Qwen3.8 as a drop-in replacement.
Token Plan access carries 2 published tiers: Lite at $6 for 2,500 credits per week, and Pro at $68 for 40,000 credits per week with 6–8 concurrent agents. Alibaba has not announced standard per-token API pricing. Its predecessor, Qwen3.7-Max, launched May 20, 2026 as an API-only model with more than 1 trillion parameters, priced at $2.50 per million input tokens and $7.50 per million output tokens, and scored 92.4% on GPQA Diamond, 80.4% on SWE-bench Verified, and 69.7% on Terminal-Bench 2.0.
Why a 2.4T Open-Weight Model Matters for Businesses
An open-weight 2.4T model would give businesses frontier-scale AI they can self-host, fine-tune without vendor permission, and run at hardware cost rather than per-token cost. Kimi K3, at 2.8 trillion parameters, is the largest open-weight model announced to date, and DeepSeek V4 Pro at 1.6 trillion is the largest whose weights are downloadable today — Qwen3.8 would land second among announced open-weight models and 50% above the currently downloadable ceiling. Companies weighing this against commercial options can compare the shipped alternatives in our guide to the 15 Best AI Tools for Business in 2026, which covers the paid AI platforms available today rather than promised releases.
One disclosure gap blocks any serious cost modelling: Alibaba has not published the active parameter count. In a sparse mixture-of-experts model, active parameters — not total parameters — determine inference compute and therefore the hardware bill. A 2.4T total with an undisclosed active count tells procurement teams nothing about what serving the model costs.
- Kimi K3 Is Here: Open-Weight Model That Rivals Claude Fable 5 — the 2.8T Moonshot AI model whose weights are scheduled for July 27
- Alibaba Bans Claude Code Over Backdoor Fears — July 10 Deadline — our coverage of Qoder, the platform now hosting the Qwen3.8 preview
- Mira Murati’s Inkling: Open-Weight AI You Can Fine-Tune Yourself — the US entrant in the same open-weight shift
What Alibaba Has Not Disclosed
Alibaba has withheld 5 details that determine whether Qwen3.8 is usable in production: the open-weight release date, the license, the active parameter count, independent benchmark results, and standard API pricing. The “second only to Fable 5” ranking is Alibaba’s own internal evaluation, and no third party has verified it.
Qwen’s release pattern gives one reference point. Qwen3.6-27B, published April 22, 2026, remains the latest open-weight general model in the family, and Alibaba shipped it after a closed API version. Alibaba also holds a 36% stake in Moonshot AI, which builds the competing Kimi K3, and Apple’s China rollout of Apple Intelligence runs on Qwen following regulatory approval on July 15, 2026. Teams that want to run whatever weights do land can see how the tooling has matured in our report on LM Studio Bionic: The AI Agent Built for Open Models.
Two trillion-parameter open-weight models announced inside one week — Kimi K3 at 2.8T and Qwen3.8 at 2.4T — mark the point where the open-weight frontier stopped trailing the closed labs by a generation. That is the real story, and it is worth separating from this specific announcement, which is a preview with a promise attached. Businesses should treat Qwen3.8 as a planning signal, not a procurement decision: the license is unknown, the serving cost is unknowable without an active parameter count, and the only performance claim on record is the vendor’s own. The date to watch is July 27, when Kimi K3’s weights are scheduled to drop — that release, not this announcement, will show what a frontier-scale open model actually costs to run.
Frequently Asked Questions
Is the Qwen 3.8 open weight model available to download?
No. Alibaba announced on July 19, 2026 that Qwen3.8 will go open-weight “soon” and has published no release date, license, or model card. As of July 22, 2026, only Qwen3.8-Max-Preview is accessible, through Alibaba’s Token Plan, Qoder, and QoderWork platforms.
How many parameters does Qwen3.8 have?
Qwen3.8 has 2.4 trillion total parameters in a sparse mixture-of-experts architecture. Alibaba has not disclosed the active parameter count, which is the figure that determines inference cost in a mixture-of-experts model.
Is Qwen3.8 better than Claude Fable 5?
No independent evidence supports that. Alibaba states Qwen3.8 ranks “second only to Fable 5” based on its internal evaluations, and no third-party benchmark results have been published for the model.
What does Qwen3.8 cost?
Alibaba prices preview access through Token Plan at $6 for 2,500 credits per week (Lite) and $68 for 40,000 credits per week (Pro, 6–8 concurrent agents). Standard per-token API pricing has not been announced.
