Commenters characterized Qwen4-Max as a capable but not clearly leading model. One cited its Artificial Analysis score of 43 and average task cost of $3.27, comparing it unfavorably with a higher-scoring, much cheaper GPT model; others called it roughly comparable to Kimi. A user’s pelican-image tests found quality generally improved at higher reasoning effort, though they preferred a Qwen 3.8 result at the highest setting. Several commenters questioned whether benchmark scores and cost estimates reflect real-world use: some argued the aggregate index changes too often or favors particular providers, while one agent-benchmark builder said Qwen 3.8 performs well when speed is excluded but loses ground because it generates tokens slowly.
Much of the interest centers on forthcoming smaller Qwen models, especially a 27B version that local users hope will run well on consumer hardware. Commenters welcomed efficiency improvements and anticipated a “Flash” model, while debating whether open-weight models are falling behind closed frontier systems. Some attributed a widening gap to Western labs’ greater compute and access to shared research; others saw recent open-model progress as evidence the gap may stabilize or close. Several users also criticized apparent political sensitivity in Qwen’s answers about China, while another suggested refusals may be imposed by the API rather than embedded in the weights. The model’s strong biotech performance drew praise, and commenters debated whether its benchmark showing reflects genuine limitations, serving speed, or evaluation choices.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.