Skip to content
Bud's Wiki

Qwen3-Max-Preview (Instruct) — Alibaba Qwen's 1T+ parameter flagship preview

TL;DR

Section titled “TL;DR”

Qwen team’s Sep 5, 2025 announcement tweet for Qwen3-Max-Preview (Instruct), described as Alibaba’s biggest model yet at over 1 trillion parameters. Distributed via Qwen Chat and the Alibaba Cloud Model Studio API (model id qwen3-max-preview), not as open weights at the preview stage. The tweet positions it as beating the prior Qwen3-235B-A22B-2507 internally, with broader knowledge, better conversation, and stronger agentic / instruction-following behavior. Framed by Qwen as evidence that “scaling works” with an official release teased to “surprise you even more.”

Key claims

Section titled “Key claims”
  • Qwen3-Max-Preview (Instruct) is announced as Qwen’s largest model yet, with over 1 trillion parameters [tweet body].
  • Availability at preview is via Qwen Chat and the Alibaba Cloud Model Studio API only — no open-weights download accompanies the preview [tweet body, embedded URLs].
  • Qwen’s internal benchmarks claim Qwen3-Max-Preview beats their previous best Qwen3-235B-A22B-2507 [tweet body]. No external benchmark table is provided in the tweet itself.
  • Qwen frames the improvements as “stronger performance, broader knowledge, better at conversations, agentic tasks & instruction following” relative to the prior best [tweet body].
  • An official (non-preview) release is teased as forthcoming and expected to exceed the preview’s numbers [tweet body].

Method

Section titled “Method”

This is a product announcement tweet — no architecture, training, or evaluation details. The only architectural signal is the parameter count (“>1 trillion”), with no statement on whether the 1T+ figure is total parameters or activated, nor whether it is dense or MoE. Given Qwen3-235B-A22B-2507 (the explicit comparison point) is a 235B-total / 22B-activated MoE, the natural reading is that Qwen3-Max-Preview is also MoE at much larger total scale — but this is inference from the comparison, not stated.

The Alibaba Cloud Model Studio link advertises the model id qwen3-max-preview for API use; the Qwen Chat link routes to chat.qwen.ai for the consumer surface.

Results

Section titled “Results”

No benchmark numbers in the tweet — only the relative claim of beating Qwen3-235B-A22B-2507 on internal evals, and a qualitative claim about agentic tasks and instruction following from “early user feedback.” Engagement at filing time: 734.5K views, 3.7K likes, 766 reposts, 244 replies.

Why it’s interesting

Section titled “Why it’s interesting”

A new datapoint in the rapidly-saturating Qwen release cadence the wiki has been tracking — companion to Qwen3.7-Max: The Agent Frontier (Qwen) (the later Qwen3.7-Max agent-frontier announcement), Qwen3.6-35B-A3B Model Card, Qwen3.6-Plus: Towards Real World Agents, and the Qwen3-VL embedding/reranker series (Qwen3-VL-Embedding-8B). Unlike most Qwen entries on the wiki, this one is API-only at preview — sits with Gemini Embedding 2 (Gemini Embedding 2: SOTA multimodal embedding model (Google product announcement)) in the closed-but-API-accessible cohort that the Open foundation-model releases page calls out as the natural baseline cohort for open frontier releases to position against. The 1T+ parameter claim — if confirmed as MoE total parameters by the eventual technical report — would put Qwen3-Max in the same scale bracket as Kimi K2 (1T total) and the unreleased GPT-5 class.

See also

Section titled “See also”