A.I.
Rycode

Jeffrey 2 min read

Open weight

Qwen3.8 Max (Alibaba)

Qwen3.8 Max is Alibaba’s lead Qwen3.8 SKU for hosted API and product surfaces. The open-weight checkpoint tied to this generation is Qwen3.8-2.4T-A95B on Hugging Face. The HF card states that hosted Qwen3.8-Max builds on that checkpoint with extra features (vision input, non-thinking support, 1M context by default, built-in tools).

Status
Current
AA score
45 (Qwen3.8 Max (0902))Artificial Analysis Intelligence Index, one independent score across many tests, fetched 22 Sep 2026
Last checked
22 Sep 2026newest source date on this page

On the Artificial Analysis Intelligence Index (fetched 22 Sep 2026 ET):

- Qwen3.8 Max (0902) scores 45 on the main leaderboard - Open checkpoint Qwen3.8 2.4T A95B scores 40 on AA’s open-source models table

Treat Max (0902) and the downloadable 2.4T checkpoint as related but not identical SKUs when you compare Index rows.

Training

Qwen materials describe Qwen3.8 as the strongest open-model generation in the Qwen family to date, with gains in coding, professional work, research, and long-horizon agents. Detailed pretrain mix and FLOPs belong in the lab blog / report when cited from primary pages.

Inference

Hosted Model Studio model id: qwen3.8-max. Snapshot qwen3.8-max-0902 (alias qwen3.8-max-2026-09-02) is documented as an upgraded Max snapshot (coding, agents, vision refinements) with a 1M context window retained.

Open weights (Qwen/Qwen3.8-2.4T-A95B): text-only on the HF card; thinking mode required; reasoning effort levels include xhigh (default), medium, and low. Serving recipes: SGLang, vLLM, TokenSpeed.

Context

  • Hosted Max (Alibaba Cloud Model Studio): context window 1,000,000 tokens (docs table)
  • Open 2.4T card: 262,144 natively, extensible up to about 1,010,000 tokens
  • AA Max (0902) row: 984k context

Methods

Value (HF open card)
TypeCausal MoE language model
Total / activated parameters2.4T / ~95B
Experts512 total; 10 routed + 1 shared activated

Hosted Max adds multimodal and product features beyond the text-only open checkpoint per the HF note.

Benchmarking

Qwen publishes large coding / agent / capability tables on the HF card (labeled under Qwen3.8-Max columns). Prefer lab primary for those harnesses. Prefer AA for Index, and keep Max (0902) 45 distinct from open 2.4T A95B 40.

Hardware and weights

  • Open weights: Qwen/Qwen3.8-2.4T-A95B
  • Blog pointer on card: qwen.ai/blog?id=qwen3.8
  • GitHub: QwenLM/Qwen3.8
  • License: custom Max-line terms on the HF card (read before redistribution). Smaller Qwen dense releases may use different licenses; do not assume Apache-2.0 for the 2.4T Max-class weights.

Access

  • Alibaba Cloud Model Studio docs: qwen3.8-max
  • Qwen Cloud / API for managed Max features
  • Self-host via HF checkpoint above

Sources