Qwen Max
by Alibaba · Model · Language & code models
mixed
Strong on web-dev and agent rankings, but behind rivals on reasoning, terminal coding and vision.
#1 search
Now shipping: 2.4T total parameters, 71B active per token (25% fewer), image and text input, 1M-token context; output price cut from $6 to $4.50 per million
AccessAPIPrice$1.425 in · $4.275 out per M tokensContext1,000K tokensNeeds~4.8 TB (a smaller build)Size2.4T parameters
Why we say so
Arena puts Qwen 4 Max #6 on WebDev and #14 on Agent Arena, and Vals ranks it #11 overall. Artificial Analysis scores it 43 on its index, behind GPT-6.1 Sol at 46, and 48% on Terminal-Bench against Kimi K3.1's 67%. Roboflow found it underwhelming on vision, and Vals found it passes only 21.3% of legal tasks despite getting 95.7% of checklist items right. Fahd Mirza's hands-on tests were very positive, but he had early access.
Versions
| Version | Released | Rank |
|---|---|---|
| Qwen 4 MaxLATEST mixed Good web-dev and big efficiency gains, but trails leaders on reasoning and coding | Nov 16 | |
| Qwen 3.8 Max strong Took #1 on Arena search, but slow at about 30 tokens/s | Jul 19 | #1 search |
Evidence
- Nov 17On a satellite image, wrote Python to locate town labels, fixed its own Sydney/Gosford mix-up, derived scale, lapse-rate math, and flagged uncertainty; took 10+ minutes
- Nov 17Scores 48% on Terminal-Bench 4.0 against Kimi K3.1 Pro's 67%; Kimi costs $3.68 per index task
- Nov 17Placed #6 in Image-to-WebDev, just above Claude Opus 5 and GPT-6 Sol
- Nov 16Qwen4-Max launched by Alibaba; commenters say it is hosted-only with no open weights yet and describe it as a 2.4T-parameter model
- Nov 16Scores 43 on the Artificial Analysis index at $3.27 average cost; GPT-6.1 Sol (medium) scores 46 at $0.36
Show all 27
- Nov 16Leads on legal agentic work, bioinformatics and long-horizon tasks, edges ahead in agentic coding, but Claude Opus 5.5 still tops many benchmarks
- Nov 16Turned a hand-drawn doner kebab machine sketch into the best interactive HTML simulation the reviewer has seen, after 15 minutes of thinking
- Nov 14Qwen 4 Max Preview shows much better token efficiency at top reasoning and better vision, with slightly lower general intelligence than 3.8 Max
- Nov 11Noticed the seatbelt strap was under the arm in a photo, which Qwen 3.8 could not infer; gave realistic per-country enforcement answers
- Nov 9#1 on Arena search (1260)
- Oct 13#2 on Arena search (1261)
- Sep 20Larger model Occamy was said to compete with on some benchmarks
- Sep 6Post-trained further on coding and co-work; same 2.4T-parameter architecture with 1M context; beats the earlier 3.8 Max, some benchmarks by over 10 points
- Sep 6Beats Opus 5 and GPT 5.6 Sol on some agentic and knowledge-work benchmarks; not open-sourced, available via Qwen Cloud API
- Sep 2Run through Hermes Agent on a Dockerized full-stack app, it found and fixed a queue-ordering bug in one targeted pass and tested it
- Aug 16Qwen3.8-Max leaves preview as Alibaba's most capable model, a 2.4T-parameter mixture of experts aimed at long-horizon autonomous work
- Aug 16Two points below Grok 4.6 on the AA index, above DeepSeek V4, near Claude and GPT-5.6; pricier than Grok, cheaper than GPT and Claude
- Aug 13Hosted platform handles video by sampling frames plus audio, not a continuous stream; tracked three people across cuts and inferred the video's theme
- Aug 9Matches Fable 5 and GPT 5.6 Sol on some agentic software-engineering benchmarks and beats Opus 4.8
- Aug 9One point below Kimi K3 on Artificial Analysis index despite about 400B fewer parameters; API slightly pricier than Kimi
- Aug 4In one-shot 3D oil-rig landing page (Three.js+GSAP), produced the best-looking, most shippable page; reviewer says it wins on taste
- Aug 4Named winner of the bug-fix agent test on time, tokens and clean cleanup; Kimi K3 close behind, no winner on correctness
- Aug 3On the WebGL landing-page prompt, judged clearly better than DeepSeek's attempt from the prior week
- Jul 26Qwen3.8-Max-Preview, a 2.4T-parameter flagship, live on Qwen Chat and the Token Plan (Coder and Coder Work)
- Jul 26Alibaba says it is second only to Fable 5; early reports suggest it is weaker than Kimi K3
- Jul 19Reviewer found its HTML canvas railway simulation (day/night, weather, steam, no collisions) better than Fable 5's output
- Jul 19Wrote original couplets in the style of local poets across about 79 languages; reviewer found it looked good, with one invented poet
Cost to run
Alternatives
Made for Qwen Max