Qwen 4 Max
The flagship large model in Alibaba Cloud's Qwen 4 series, positioned as an open-source model with ultra-long context and multimodal capabilities. It supports a context window of up to one million tokens, image input, and tool calling.
Core Facts
The Qwen4 Max Preview runs until November 25.
Qwen2.5-Max-0902 supports a context window of 1 million tokens.
Modern reasoning models allow users to set thinking intensity levels ranging from low to high and then x-high; the higher the level, the more time the model spends on internal reasoning before generating a response.
Timeline (last 90 days)
Qwen 4 Max is better suited as an auxiliary tool for complex projects rather than the first choice for lightweight rapid prototyping.
The generated game prototype is excellent in visual presentation (scene modeling, lighting effects) but incomplete in gameplay logic, such as no health display, no minimap, no running, and AI enemy behavior that is not intelligent enough.
The game project generated by Qwen 4 Max is named Blacksite, a first-person shooter prototype, and the model itself considers it not yet on par with Call of Duty.
In the OpenCode terminal environment, Qwen 4 Max took several hours to complete a Call of Duty-style first-person shooter development test, going through 4 rounds of iteration, consuming 360,000 tokens at a cost of about $31.
Compared with the previous generation Qwen 3.8 Max, Qwen 4 Max requires substantially fewer tokens to complete the same type of task.
Qwen 4 Max is a natively multimodal architecture that can directly understand image content.
A single generation by Qwen 4 Max struggles to deliver a complete, playable result and requires manual intervention or additional prompts to fill in the details.
Alibaba's Qwen 4 Max officially launched on November 11, and the company released no benchmark results or model card.
The Qwen 4 Max preview ends on November 25, after which the model will be retired.
Tester Ajay measured Qwen 4 Max's generation speed at about 47 tokens per second, and a river scene took 30 minutes.
40 more timeline events
All Facts (20)
The Qwen4 Max Preview runs until November 25.
70%VerifiedQwen2.5-Max-0902 supports a context window of 1 million tokens.
70%VerifiedModern reasoning models allow users to set thinking intensity levels ranging from low to high and then x-high; the higher the level, the more time the model spends on internal reasoning before generating a response.
70%VerifiedBoth Qwen 3 Max and DeepSeek-V3 adopt the MoE architecture with total parameter counts reaching hundreds of billions, but the parameters actually activated during inference are only a small fraction of the total.
70%VerifiedThe reviewer believes that Qwen3.8 27B is stronger than the previous generation Qwen3.7 in almost all aspects of coding capability.
65%VerifiedArtificial Analysis gave Qwen 3.8 an intelligence index of 52, tying GPT at the full reasoning tier at 5.6, with an agent score of 51, higher than Claude Opus 4.8.
65%Partially VerifiedCurrently Qwen 4 Max is available only in preview form, with no open weights, and model files cannot be downloaded.
65%Partially VerifiedAs a preview version, Qwen 4 Max's performance fluctuates considerably, so it is recommended to try it first on non-critical paths and do a full evaluation after the official version is released.
65%Partially VerifiedQwen4-Max is positioned as a reasoning model, natively supports visual input, and has a context window of 1 million tokens.
65%UnverifiedThe Qwen 4 Max preview ends on November 25, after which the model will be retired.
60%UnverifiedThe Qwen 4 Max preview cannot be downloaded and has no token-billed API; the only entry point is Alibaba Cloud's token package, billed by monthly subscription.
60%UnverifiedAcross five hands-on coding tasks by a YouTube creator, Qwen 4 Max's total cost was about 1.65 + 0.64 + 1.26 + 0.29 USD (the cost of the fifth task was not explicitly listed), with per-task cost far below the compared Claude models.
60%UnverifiedOn the DeepSWE coding benchmark, Qwen 4 Max scored 75.7, ranking first, but Kimi K3.1 scored 75.5 and GPT 6.1 Sol scored 75.2, so the gap is extremely small.
60%UnverifiedOverall, Qwen 4 Max leads in 5 of 16 benchmarks.
60%UnverifiedQwen4-Max offers two reasoning modes, thinking mode and standard mode; thinking mode generates additional reasoning steps before outputting the answer, suiting complex tasks, but the reasoning tokens count toward the output budget, increasing processing time and API costs.
60%UnverifiedUsers can try Qwen 4 Max for free on QwenChat, and the API subscription "Token Plan" starts at about $6 per month, or $3 for new users.
60%UnverifiedIn the Atomic Agent team test, Qwen 4 Max passed all 59 hidden tests in three out of three runs, while the previous generation Qwen 3.8 Max failed to complete even once; Qwen 4 Max took about 25 minutes, Opus 5.5 about 4 minutes, and GPT-6.1 Sol less than two and a half minutes, and Sol cost about 29 cents while Qwen cost about 65 cents.
60%UnverifiedIn Alibaba's benchmarks, Qwen 4 Max won with 75.7 on the DeepSUE coding agent test, led with 46.7 on Agent's Last Exam, and scored 21.3 on HLAB (Opus 5.5 was 9.2); however, it scored 48.5 on Terminal Bench 4.0 (Opus 5.5 was 59.6, Kimi was 67.2), 31.5 on Program Bench (Opus 5.5 was 65), and 51.7 on Humanity's Last Exam (Opus 5.5 was 59.8). In total, Qwen 4 Max led in 3 rows and Opus 5.5 led in 7 rows.
60%UnverifiedQwen 4 Max is a natively multimodal architecture that can directly understand image content.
50%UnverifiedCompared with the previous generation Qwen 3.8 Max, Qwen 4 Max requires substantially fewer tokens to complete the same type of task.
50%