Magic Tools
Back to all briefs

Dev Breakfast · 2026-11-17

Today's headline: Qwen4-Max Releases Task-Cost Benchmark Results.

November 17, 20263 min readDev Breakfast
🍳 Today's Headlinethe one deep dive of the day

Qwen4-Max Releases Task-Cost Benchmark Results

In an introduction published on August 3, the Qwen team unveiled Qwen4-Max: a new architecture with native vision support and a 1 million-token context window. According to official cost figures, the average cost per task in its in-house Work360 benchmark fell from $4.43 to $1.67, a reduction of about 62%, while its score rose from the previous-generation Qwen3.8-Max's 47.74 to 52.69. This is a reduction in task costs within a specific benchmark, not an across-the-board discount of the same size on API unit prices.

The direct comparison for this upgrade is Qwen3.8-Max. The team changed both the model architecture and API pricing, and said that cached-input prices fell particularly sharply, though the article did not list specific unit prices. Usage is also declining: on Work360, token usage at the Xhigh and High reasoning effort levels fell by about 24% and 35%, respectively; on Terminal-Bench 4.0, the reductions were about 48% and 49%. The 62% figure therefore reflects the combined effects of usage and pricing, and cannot be applied directly to your coding-task bills.

The two benchmarks also cover different kinds of work. The Work360 taxonomy that the Qwen team established for professional work includes 26 primary domains, 182 secondary domains, and 721 tertiary specialty domains, forming the basis for the benchmark; it is not a standalone programming test. The same introduction also demonstrated Blender modeling and game development using Unity, Unreal Engine, and Godot. The workflow described by the team goes beyond generating code to include running games, reading visual and interactive feedback, and then adjusting cameras, animation, lighting, and more. These demonstrations illustrate what the team describes as long-task execution capabilities, but cannot replace data on success rates for complete project delivery.

The parameter count also needs to be broken down. The stated 2.4 trillion parameters refer to the backbone, with an additional 328 billion N-gram Embedding parameters. Each token activates 71 billion parameters, rather than all parameters participating in every computation. The long-context mechanism pairs each QSA layer with three Gated DeltaNet layers: the latter compress history into a fixed-size state, while QSA uses a lightweight indexer to select relevant context blocks. The team attributes the efficiency gains to this kind of sparse computation and architectural changes; a 1 million-token window capacity alone does not mean the model can use every detail within that window with equal accuracy.

What has been released so far consists of official benchmarks and demonstrations; the article does not provide a complete price list and usage breakdown sufficient to recalculate task costs. The base-model evaluation also has limitations: in internal tests across four domain categories, Qwen4-Max-Base reduced BPB (bits per byte, where lower is better) by about 5%–13% compared with its predecessor, while reducing active parameters by 25%. It had the lowest BPB among the five model checkpoints tested, but the team specified that GLM-5.3 and Kimi-K3.1-Pro in the comparison had undergone post-training, which could affect comparisons with a base model.

Sources:


This article selects one story for a closer look from 60 items on X / Hacker News / GitHub Trending over the past 24 hours. The content was generated with LLM assistance and fact-checked, with background drawing in part on this site's previous coverage. Original source links are included at the end; cross-check before making important decisions.

Like this brief? Get it by email

Daily AI coding picks at 8:00, plus a hands-on field-notes issue every Saturday. Written in Chinese.

This page is auto-generated by LLM aggregation; please cross-check with original sources.

Dev Breakfast · Qwen4-Max Releases Task-Cost Benchmark Results