Skip to content
Qwen4-Max Cuts Task Cost 62% While Boosting Scores: Architecture & Benchmarks

Qwen4-Max Cuts Task Cost 62% While Boosting Scores: Architecture & Benchmarks

Tongyi's Qwen4-Max reduces per-task cost from $4.43 to $1.67 (62% less) on its Work360 benchmark while raising scores from 47.74 to 52.69, using a 2.4T-parameter architecture with 710M activated parameters, phrase embeddings, gated recurrent layers, mixture-of-experts, and sparse attention, with training loss declining steadily without spikes.

AI Frontier Lectures
AI Frontier Lectures
AI Frontier Lectures

On November 16, the Tongyi team released Qwen4-Max. According to the official blog, on their internal Work360 benchmark the average cost to complete a task dropped from $4.43 with the previous Qwen3.8-Max to $1.67, a 62% reduction.

The blog breaks the improvement into two factors. First, token output decreased: on Work360 the high tier emits roughly 35% fewer tokens and the ultra-high tier about 24% fewer. On the separate Terminal-Bench 4.0 suite the reductions are approximately 49% and 48% respectively, with the high tier again leading. Second, API pricing was cut, most notably for cached input tokens; the new per-million-token rates are not printed in the blog. On the same Work360 benchmark the task score rose from 47.74 to 52.69.

Work360 is Tongyi's proprietary office benchmark comprising 26 top-level categories, 182 sub-categories, and 721 directions. Both cost and score figures are measured on this suite, not on a public leaderboard. The blog does not detail how the $4.43 figure is derived from per-token pricing and token counts.

Work360 cost and token reduction table comparing Qwen4-Max vs Qwen3.8-Max
Work360 cost and token reduction table comparing Qwen4-Max vs Qwen3.8-Max

The table above, reconstructed from the blog's numbers, shows four rows each comparing Qwen4-Max against the previous generation. Cost and score are on Work360; the two token rows show the high tier down ~35% and the ultra-high tier down ~24%. Token savings on Terminal-Bench are not included in this table.

Scatter plots of tokens per task vs score for Work360 and Terminal-Bench 4.0
Scatter plots of tokens per task vs score for Work360 and Terminal-Bench 4.0

The chart shows two panels: Work360 (left) and Terminal-Bench 4.0 (right). The horizontal axis is tokens per task, the vertical axis is score. The solid dark-blue line represents Qwen4-Max; the dashed light-blue line is Qwen3.8-Max-0902. Three operating points (Low, High, Xhigh) are plotted for each model. In both panels the new model's curve lies to the left (fewer tokens) and higher (better score) than the old one. The percentage reductions cited in the text are not marked on the points themselves.

Bar chart comparing Qwen4-Max against Opus, GPT variants, and Kimi across multiple benchmarks
Bar chart comparing Qwen4-Max against Opus, GPT variants, and Kimi across multiple benchmarks

A 16-bar chart compares Qwen4-Max (dark blue) against the previous generation (light blue) and several external models (gray bars: Opus, two GPT tiers, Kimi) across multiple benchmarks. This chart compares scores only, not per-task cost.

Model Architecture

Qwen4-Max uses a new architecture for the Qwen4 generation. The backbone contains 2.4 trillion total parameters, with 710 million activated per token. An additional 3.28 billion parameters are allocated to phrase embeddings. The context window is 1 million tokens and the model accepts image inputs.

Architecture diagram: vocabulary, phrase embedding, three gated recurrent layers with residuals, MoE, sparse attention, prediction head, multi-token prediction
Architecture diagram: vocabulary, phrase embedding, three gated recurrent layers with residuals, MoE, sparse attention, prediction head, multi-token prediction

The diagram reads bottom-up. Input passes through the vocabulary layer, then a second layer queries the phrase-embedding block (only at this layer). Above that sit three gated recurrent layers; each layer has four residual streams with orange blocks for reads and blue blocks for writes. A mixture-of-experts module sits to the right. Higher up is a sparse-attention layer with the same structural pattern. At the top are the prediction head and a multi-token prediction module.

The blog states that pre-training loss decreased monotonically without spikes, attributing this stability to the architecture and training regimen. The curve is reproduced below.

Training loss curve: steep early drop then gradual decline, no spikes
Training loss curve: steep early drop then gradual decline, no spikes

The horizontal axis shows training progress from 0% to 100%; the vertical axis is training loss. A single dark-cyan curve drops sharply at the start and continues a smooth decline through the second half, with no intermediate spikes.

Compression-Rate Evaluation

Internal evaluation used compression rate (bits per byte, lower is better) on four domains: knowledge, code, Q&A, and STEM. This is a separate test set from Work360.

Compression rate bars: Qwen4-Max-Base (71B/2.4T) vs previous base (95B/2.4T) and post-trained GLM/Kimi
Compression rate bars: Qwen4-Max-Base (71B/2.4T) vs previous base (95B/2.4T) and post-trained GLM/Kimi

Each panel's vertical axis is 10× bits per byte. The rightmost blue bar is Qwen4-Max-Base labeled 71B/2.4T; the leftmost light bar is the previous-generation base labeled 95B/2.4T. The blog reports that the blue bar is lower on all four domains by roughly 5% to 13%, while activating 25% fewer parameters per token. The caption notes that GLM and Kimi bars are post-training checkpoints and are not directly comparable to the base models.

Availability

Qwen4-Max can be accessed via QwenCloud's pay-as-you-go API or Token Plan, and inside QwenWork, Qoder, and QwenStudio. Model weights have not been released. A companion plugin, Qwen-Live-Delivery-Plugin, is publicly available on GitHub.

Source: Tongyi official blog, November 16: https://qwen.ai/blog?id=qwen4-max Long-task plugin: https://github.com/QwenLM/Qwen-Live-Delivery-Plugin

Code example

👀点击关注
AI前沿快速Follow
Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

mixture of expertslarge language modelbenchmarkmodel architecturecost reductiontraining stabilityQwen4-Maxcompression rate
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.