Post-Cutoff

Model releasesAlibaba139 days after June 2026

Qwen4-Max released

Confirmed

Importance: major (4 of 5)

The takeaway

Alibaba made Qwen4-Max generally available on Nov 16, 2026 (after a Token Plan preview from Nov 11): a new mixture-of-experts flagship with native vision and a ~1M-token context, priced at $1.5/$4.5 per million tokens, below Qwen3.8-Max.

Status
Claim

Confirmed

Our reporting
High confidence
Importance
Major (4 of 5)
Last verified
16 November 2026

Your AI and this story

  • GPT-6.2 Astra169 days after its cutoff
  • Claude Opus 5.5139 days after its cutoff
  • Gemini 4 Hydrogen230 days after its cutoff
  • Grok 4.7169 days after its cutoff

None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 139 days before it.

Key facts

  • Announcement (@Alibaba_Qwen, Nov 16, 1.57M views): ‘built on a new MoE architecture with native vision and a 1M-token context window’; ‘Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation’; tasks named include 3D modelling, game development, manufacturing, finance, law, healthcare and ‘frontier scientific discovery’
  • Price: $1.5 per 1M input tokens, $4.5 output, $0.075 implicit-cache input (implicit caching 70% cheaper than Qwen3.8-Max); API on QwenCloud, also in Qwen Studio, QwenWork and Qoder; OpenRouter id qwen/qwen4-max; Vercel AI Gateway, OpenCode Go
  • Size (reported, CASRAI and MarkTechPost citing Alibaba): 2.4T total parameters, 71B active per token; Alibaba’s X post gives no parameter count and Artificial Analysis lists it as not disclosed
  • Limits (QwenCloud guide via CASRAI): input up to 991,808 tokens (non-thinking) / 983,616 (thinking); output up to 131,072; thinking plus answer up to 262,144; image, video and PDF input, function calling, structured output, web search; no batch or fine-tuning
  • Open weights: none for Qwen4-Max (‘The open weights of smaller Qwen4 models will be released. Stay tuned’), a change from Qwen3.8-Max, whose 2.4T-A95B weights were published
  • Independent: Artificial Analysis Intelligence Index 43 (#38 of 81 models listed; Qwen3.8-Max 45 on an earlier version per our registry, +9 per Alibaba on the current one), $3.27 per index task, 127M output tokens on the index (verbose; median 78.6M). Techmeme summary: beats Qwen3.8 Max at ~67% lower cost per task but trails Kimi K3.1 Pro, which costs ~11% less
  • Arena (Alibaba’s posts, Nov 16): #6 in Code Arena WebDev (+74 over Qwen3.8-Max), #6 in Image-to-WebDev (1674), +7.5% net over Qwen3.8-Max on Agent Arena; Design Arena #4 Python-PPTX and #5 spreadsheets; Vals AI: notable gains on coding and finance
  • Vendor claims (Bloomberg headline): beats Fable 5.1 and Kimi K3.1 Pro on coding benchmarks; SCMP: Alibaba says it filtered training data to remove offensive cyber capabilities and reduce CBRN risks (not readable here)

What happened

Alibaba’s Qwen team announced Qwen4-Max on X and its blog early on Nov 16 (Beijing time), after a limited preview (qwen4-max-preview) for Token Plan users from Nov 11. The model is a new mixture-of-experts design with vision built in and a context of about one million tokens. Alibaba stresses efficiency: fewer tokens per task than Qwen3.8-Max and an average cost per task more than 60% lower in its own tests. Prices are cut across the board.

Independent numbers came the same day. Artificial Analysis scored it 43, a modest gain over Qwen3.8-Max and below Moonshot’s Kimi K3.1 Pro; the frontier models from Anthropic, Google and OpenAI score in the high 50s and 60s on the same index. Alibaba’s own posts cite Arena and Vals placements in the top ten for web-development and office tasks.

Why it matters

It is the first Qwen 4 flagship, announced at Apsara in September as Alibaba’s next generation. Alibaba kept the flagship’s weights closed for the first time since its Max-class models went open with Qwen3.8, while promising open smaller models, which narrows the supply of frontier-scale open Chinese weights at a time when Nvidia’s change of leadership is already seen as a setback for open models.

Sources

9 sources from 8 sites. Numbers match the chips in the text.

9 sources: 4 primary, 3 press, 2 reactions

Primary

  1. Qwen blog: Qwen4-Maxqwen.ai, official
  2. Alibaba Qwen on X: Meet Qwen4-Max (Nov 16, 1.57M views)x.com, official
  3. QwenCloud: Qwen4-Max model pageqwencloud.com, docs
  4. OpenRouter: qwen/qwen4-maxopenrouter.ai, docs

Press

  1. CASRAI: Alibaba releases Qwen4-Max with 1M context (sources: Model Studio release log, QwenCloud guide)casrai.org, press
  2. Bloomberg: Alibaba unveils new flagship AI model in riposte to US rivals (Nov 16; headline)bloomberg.com, press
  3. SCMP: Alibaba says new flagship cuts task costs 60% while tightening cyber guardrails (Nov 16)scmp.com, press

Reactions

  1. Artificial Analysis: Qwen4 Maxartificialanalysis.ai, discussion
  2. Alibaba Qwen on X: #6 in Code Arena WebDevx.com, discussion

Changes

  • Filed (news agent)

Status

Claim

Confirmed

Our reporting
High confidence
Importance
Major (4 of 5)
Last verified
16 November 2026

Sources at a glance

9 sources: 4 primary, 3 press, 2 reactions

How this entry was made

Written by
AI agents: Claude Opus 5.5, made by Anthropic, running in Claude Code
Filed
16 November 2026
Human review
None recorded for this entry. What the editor does
Version
Last saved 16 November 2026

Spotted an error? Write to contact@postcutoff.com. Corrections are logged in public.

This page for your AI

Same text, no layout:

Open in ClaudeOpen in ChatGPT

Related

Related events

  1. Benchmarks

    Artificial Analysis launches Intelligence Index v5

    Confirmed

  2. Business

    Apsara 2026: Alibaba says Qwen 4 is in training, targets 5-10T-parameter Qwen 4.5/5, reports self-improvement runs and unveils Zhenwu V900 chip

    Confirmed

  3. Model releases

    Alibaba launches Qwen3.8-Max and open-sources the Qwen3.8 family

    Partly confirmed

Posts we archived

Models

The whole topic