|

Alibaba Previews Qwen3.8-Max, a 2.4 Trillion-Parameter Multimodal Model, Days After Moonshot’s Kimi K3 Open-Weight Launch

ByRicardo

On July 19, Alibaba’s Qwen staff previewed Qwen3.8-Max-Preview, the subsequent flagship within the Qwen household. The analysis staff describes it as a 2.4 trillion-parameter mannequin, ‘second solely to Fable 5’ among the many programs it benchmarked. The preview is stay now. The benchmark desk, mannequin card, and license should not.

The July nineteenth 2026 announcement landed throughout the World AI Conference (WAIC) in Shanghai. It additionally arrived two days after Moonshot AI launched Kimi K3, a 2.8 trillion-parameter open-weight mannequin. The timing is the story as a lot because the mannequin.

This article separates what Alibaba confirmed from what it solely claimed. Every efficiency determine under carries that caveat.

What Qwen introduced

The Qwen account posted that Qwen3.8 is launching and going open-weight quickly. It referred to as the mannequin ‘some of the highly effective out there at the moment, similar to main frontier programs.

The preview construct is actual and purchasable. Access runs by means of Alibaba’s Token Plan subscription. The preview is obtainable at 10% of ordinary pricing.

Qwen developer Shuai Bai added technical detail. He described Qwen3.8 because the staff’s first multimodal mannequin above 1 trillion parameters. It processes textual content, photographs, video, and paperwork. Alibaba staff states the mannequin ought to beat Qwen3.7-Max on coding, full-stack growth, information evaluation, and workplace workflows.

Interactive Explainer



Qwen3.8 Explainer

Marktechpost Interactive Explainer

Qwen3.8-Max-Preview: what Alibaba confirmed, what it solely claimed

A 2.4T-parameter multimodal preview shipped earlier than any benchmark, mannequin card, or license. Explore the details under.

Confirmed vs Claimed
Parameter scale
Serving-cost calculator
Verified 3.7-Max baseline

✓
Preview is stay and purchasable. Qwen3.8-Max-Preview is offered by means of Alibaba Token Plan, Qoder, and QoderWork at 10% of ordinary pricing.
✓
Sparse Mixture-of-Experts, multimodal. Developer Shuai Bai says it’s the staff’s first multimodal mannequin above 1T parameters, dealing with photographs, video, and paperwork.
✓
OpenAI and Anthropic protocol compatibility. Existing coding brokers can level at Qwen3.8 with out rebuilding their harnesses.
✕
2.4 trillion parameters. This is Alibaba’s personal determine. No mannequin card or specification confirms it.
✕
“Second solely to Fable 5.” No benchmark desk has been revealed. The rating rests on inside analysis.
✕
Open weights “quickly.” No date, no license, no Hugging Face repository. Alibaba’s final two Max flagships shipped closed.
✕
Active parameters per token. Undisclosed — the only quantity that decides actual serving value for a sparse MoE mannequin.

Status as of July 19, 2026. Toggle a truth kind utilizing the tabs above.

Total parameters, publicly disclosed frontier fashions, July 2026. Qwen3.8’s 2.4T is Alibaba’s declare, not a verified determine.

Total rely just isn’t usable compute. A sparse MoE mannequin prompts solely a fraction of those parameters per token.

If the total 2.4T mannequin shipped open-weight, what would it not take to load?


—
Weights solely
—
Nvidia H200 (141GB) playing cards

Estimate for weights alone; add roughly 20–30% for the KV cache and runtime overhead. Active-parameter serving can be far cheaper, however Alibaba has not revealed that quantity.

The verified predecessor. Every revealed Qwen3.8 “functionality” quantity is basically a Qwen3.7-Max quantity till Alibaba releases a benchmark desk.
92.4
GPQA Diamond
80.4%
SWE-bench Verified
69.7
Terminal-Bench 2.0
1M
Context window (tokens)
$1.25
Per 1M enter tokens
$3.75
Per 1M output tokens

Qwen3.7-Max, May 2026, closed weights. Alibaba’s historic edge has been price-to-performance, not topping a leaderboard.

Marktechpost
Data verified July 19, 2026 · Figures marked claimed are Alibaba’s personal

Similar Posts