Alibaba just open-sourced a 2.4 trillion parameter AI model — good luck ever running it yourself

it beats Claude on some benchmarks and rivals GPT on others, and it's free to download next week. here's the hardware math that ruins the fun.

Aliteq
Ravi Malhotra · Hardware Editor

Key point

Qwen3.8-Max is a 2.4-trillion-parameter mixture-of-experts model, Alibaba's largest release, launched August 3, 2026.

Key point

It's the first Qwen-Max-class model ever to get an open-weights release — previous Max-tier models stayed closed.

Key point

Scores: 93.0 PaperBench, 92.6 GPQA Diamond, 86.6 Terminal-Bench 2.1, 82.8 IFBench — competitive with, but not uniformly ahead of, GPT-5.6 Sol and Claude Opus 4.8/Fable 5.

Key point

Weights land on Hugging Face and ModelScope next week; a hosted API is live today at $2/M input and $6/M output tokens.

Key point

Even quantized to 4-bit, the raw weights need roughly 1.2 terabytes of memory — nowhere near a consumer GPU, or two.

Who this actually matters to

This is a genuine win for cloud infrastructure teams, researchers with real GPU cluster access, and anyone Alibaba can pull off a competitor's API — not for someone with a single high-end card and a…

Aliteq

Read the full story

Alibaba just open-sourced a 2.4 trillion parameter AI model — good luck ever running it yourself

 Read the full story on Aliteq