Alibaba launches Qwen4‑Max: cheaper AI work but mixed benchmark results
Alibaba unveiled Qwen4‑Max, a multimodal model that cuts task costs by about 60% while offering up to one‑million token context.
Source: VentureBeat — AIAlibaba introduced Qwen4‑Max, its newest flagship AI model, positioning it as a cost‑effective solution for enterprises that need multimodal capabilities. The model accepts text, images, and video, can call external tools, and supports up to one million tokens of context – a practical upgrade over its predecessor, Qwen3.8‑Max.
What is Qwen4‑Max and why does it matter?
According to VentureBeat, Qwen4‑Max is a proprietary model that builds on the 2.4‑trillion‑parameter backbone of the Qwen family while activating only a fraction of those parameters per token. Alibaba claims the new model improves the internal Work360 benchmark score from 47.74 to 52.69, a relative gain of about 10 %. More importantly for businesses, the company reports that the average cost of completing a task in Work360 fell from $4.43 to $1.67 – a reduction of roughly 62 %.
How does its pricing compare to rival models?
The QwenCloud API lists a price of $1.50 per million input tokens and $4.50 per million output tokens, with a cached‑input rate of $0.075. Those rates are lower than the standard prices for Anthropic’s Claude Opus 5.5 ($4.00 in, $20.00 out) and comparable to Google’s Gemini 4 Hydrogen ($1.50 in, $7.50 out). A side‑by‑side snapshot from the article shows:
| Model | Input ($/M) | Output ($/M) |
|---|---|---|
| Qwen4‑Max | $1.50 | $4.50 |
| Claude Opus 5.5 | $4.00 | $20.00 |
| Claude Sonnet 5.5 | $2.00 | $10.00 |
| GPT‑6.1 Sol | $2.00 | $10.00 |
| Gemini 4 Carbon | $2.00 | $10.00 |
| Gemini 4 Hydrogen | $1.50 | $7.50 |
| Kimi K3.1 Pro | $2.00 | $10.00 |
While Qwen4‑Max’s output cost is 55 % lower than the $10 baseline used by many competitors, Anthropic’s Haiku 5.5 can be cheaper for short prompts (as low as $0.10 in, $0.50 out for under 100 k tokens). The model also charges $1.875 per million tokens to create a five‑minute cache and $3 for a one‑hour cache, with reads at $0.075. Those extra fees mean the headline $1.67 per task figure reflects both token usage and the base pricing, not a universal discount.
How does it perform on real‑world benchmarks?
VentureBeat reproduces Alibaba’s own benchmark table, which pits Qwen4‑Max against leading U.S. models on a variety of agent‑style tasks. The model narrowly leads on the DeepSWE v1.1 coding‑agent test (75.7 vs. 69.3 for Qwen3.8‑Max) and comes close to Claude Opus 5.5 on the APEX‑Agents professional‑services suite. However, the gaps are stark in other areas:
- Terminal‑Bench 4.0 – Qwen4‑Max scores 48.5, well behind Kimi K3.1 Pro (67.2) and Opus 5.5 (59.6).
- ProgramBench – a 31.5 score versus Opus 5.5’s 65.0, indicating weaker software‑replication ability.
- ChartoGraphy – 54.2, lagging behind Opus 5.5 (66.3) and Gemini 4 Carbon (67.8).
An independent aggregate Intelligence Index places Qwen4‑Max at 43, behind Kimi K3.1 Pro (46), Claude Sonnet 5.5 (55), and the leading Claude Fable 5.2 (65). Speed is also a concern: the model reportedly takes about 32 minutes per Intelligence Index task and 159 seconds to emit the first token, situating it on the slower end of the spectrum.
What architectural changes enable token efficiency?
The model’s mixture‑of‑experts design activates only 71 billion of the 2.4‑trillion‑parameter backbone for each token, with an additional 328 billion parameters handling local token‑pattern representations. Alibaba says this results in roughly 25 % fewer active parameters than Qwen3.8‑Max. The architecture also features compressed history and selective attention aimed at making long contexts more economical – a key factor in achieving the reported 24 % token reduction at the "Xhigh" reasoning setting and 35 % at "High" within Work360.
These efficiencies translate into the near‑50 % token drop observed on the Terminal‑Bench tests, though the 62 % cost reduction combines token savings with the lower per‑token pricing.
Is the model available for on‑premise use?
Qwen4‑Max is released as a proprietary service. Alibaba provides access through the QwenCloud API, as well as integrated UI experiences in Qwen Studio, QwenWork, and Qoder. Unlike earlier Qwen releases, no downloadable weights have been announced for the Max variant, and the company has offered no concrete timeline for open‑weight versions of the larger model. Smaller Qwen4 variants are promised to be open‑source, but dates and licenses remain unspecified.
For enterprises, this means the model can be consumed via cloud API or embedded in Alibaba’s own tooling, but organizations that require on‑premise deployment or full model transparency will need to evaluate alternative providers.
How can Grafy users experiment with Qwen4‑Max?
Grafy’s Chat app lets creators and developers converse with multiple models side‑by‑side, including proprietary APIs that can be added via custom endpoints. Users interested in testing Qwen4‑Max’s cost and capability profile can configure a QwenCloud endpoint in Grafy Chat, run prompt‑to‑response experiments, and compare token usage against other models available in the platform. This approach lets teams gauge whether the lower per‑token rates translate into real‑world savings for their specific workloads without leaving the Grafy workspace.
To explore this yourself, open Grafy Chat: chat with many AI models side by side.
Frequently asked questions
Does Qwen4‑Max support native audio or video generation?
No. The model accepts text, image, and video inputs and returns text output only. Generating audio or video requires invoking external tools, such as Blender for 3D work or WonderClip for video production, as described by Alibaba.
Are the cost savings reported by Alibaba universal for all workloads?
The $1.67 average task cost reflects Alibaba’s internal Work360 benchmark, which combines token reductions and the model’s pricing structure. Independent tests (e.g., Artificial Analysis) show an 11 % cost advantage over Kimi K3.1 Pro on a different task suite, but savings will vary depending on the complexity, token length, and cache usage of each workload.
Will Alibaba eventually release the weights for Qwen4‑Max?
As of the November 2026 announcement, Qwen4‑Max is proprietary with no public commitment to release its weights. Alibaba has pledged open releases for smaller Qwen4 models, but the timeline and licensing terms for the Max variant remain undefined.


