DaisyRose-27B: 1,200-task RL took Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1
Alibaba has unveiled Qwen4-Max, a 2.4-trillion-parameter flagship model, asserting it cuts average costs for professional tasks by 62% compared to its predecessor. The company released a blog post detailing Work360 benchmarks, showing Qwen4-Max outperforming Qwen3.8-Max while using 35% fewer tokens in high-effort reasoning mode, leading to a 62% reduction in average task costs from $4.43 to $1.67.
The model maintains a 2.4-trillion-parameter mixture-of-experts backbone but activates 71 billion parameters per token, and supports up to a 1-million-token context window. Alibaba attributed the cost savings to lower token usage and revised pricing, with input tokens at $1.50 and output tokens at $4.50 per million, down from $2 and $6 respectively.
Experts advise treating Alibaba's figures as vendor benchmarks rather than guaranteed savings, emphasizing successful task completion costs instead. Alibaba also highlighted improvements in coding and agentic evaluations, including Code Arena and Agents' Last Exam, and demonstrated a 24-minute reduction in manual workpiece programming time for automotive manufacturing.
Concerning security, the model filters training data to include defensive cybersecurity tasks while removing offensive ones, improving defensive capabilities by 9.8% on the Last Secure Code Bench. However, security experts question the reliability of separating offensive and defensive knowledge through training-data filtering, recommending independent evaluations by government or external red teams.
Alibaba also emphasized data governance risks, suggesting enterprises consider prompt retention policies and jurisdictional considerations for sensitive workloads.
Written by urgent.news from Computerworld's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
This story
This is one outlet's version. Read the fullest account.
- Alibaba touts 62% lower task costs with Qwen4-Max, adds cyber safeguards computerworld.com
- Qwen4-Max scores 43 on the Artificial Analysis Intelligence Index, beating Qwen3.8 Max at ~67% lower cost per task but trailing Kimi K3.1 Pro at ~11% lower cost (Artificial Analysis) artificialanalysis.ai
- Alibaba says Qwen4-Max, which isn't open weight, beats Fable 5.1 and Kimi K3.1 Pro in coding benchmarks, and plans to release weights for smaller Qwen4 models (Luz Ding/Bloomberg) bloomberg.com
- Alibaba says it filtered Qwen4-Max's training data to remove offensive cyber capabilities and mitigate chemical, biological, radiological, and nuclear risks (Ann Cao/South China Morning Post) scmp.com
- Alibaba releases Qwen4-Max, its new flagship 2.4T-parameter MoE model with a 1M-token context window, saying it beats Qwen3.8-Max with up to ~50% fewer tokens (Qwen) qwen.ai