Alibaba releases Qwen4-Max, a 2.4T-parameter MoE model with a new hybrid architecture, 1M context and lower per-task costs than its predecessor
Edited by Elliot Marsh, Editor · written with AI from the linked sources (how)
Alibaba launched Qwen4-Max, a mixture-of-experts model (2.4T total, 71B active parameters) with native image input and a 1M-token context. It uses a hybrid attention architecture that is cheaper on long inputs. API pricing is $1.50 per million input tokens, $4.50 per million output tokens and $0.075 for cache hits. Artificial Analysis scores it 43 on its Intelligence Index, second among Chinese models behind Kimi K3.1 Pro (46). Its cost per index task is $3.27, against $10.02 for Qwen3.8-Max and $3.68 for Kimi K3.1 Pro. Alibaba also introduced a "Live Delivery" feature, which lets users inspect and steer running agents, and open-sourced the Qwen-Live-Delivery-Plugin.
Entities: Alibaba, Qwen4-Max, Qwen3.8-Max, Kimi K3.1 Pro, Opus 5.5, GPT-6
0 primary
What happened
Alibaba released Qwen4-Max, a mixture-of-experts model with 2.4T total and 71B active parameters, native image input and a 1M-token context. It uses a hybrid attention design that is cheaper on long inputs. API pricing is $1.50 per million input tokens, $4.50 per million output tokens and $0.075 for cache hits. Artificial Analysis scores it 43 on its Intelligence Index, second among Chinese models behind Kimi K3.1 Pro (46). Its cost to run the index tasks is $3.27, against $10.02 for Qwen3.8-Max and $3.68 for Kimi K3.1 Pro. Alibaba also added a "Live Delivery" feature for inspecting and steering running agents, and open-sourced the Qwen-Live-Delivery-Plugin.
Why it matters
The most useful fact is cost. Qwen4-Max is about a third of the previous Qwen flagship's cost per index task and roughly level with Kimi K3.1 Pro, so teams running long-context or agent workloads have a cheaper option to test. It is not the top performer: it trails Kimi by three index points, and the coverage gives no independent score for Opus 5.5 or GPT-6. The open-sourced plugin matters only if the agent-steering feature works as described, which is untested. Buyers should treat this as a candidate for their own trials, not a reason to switch.
What is noise
The headline "takes on Opus 5.5, GPT-6 and Kimi K3.1 Pro" is framing. The one independent benchmark puts Qwen4-Max behind Kimi, and the article itself says external results are mixed. The 62% cost cut on Work360 comes from Alibaba, not a third party. Claims of leading agent and work performance, and of suitability for industry, science and finance, are marketing until independent tests back them up. The articles also leave out how reliable the model is in real use, its latency and its rate limits.
Watch next
- 01Independent evaluations on your own workloads, plus further third-party benchmarks (Artificial Analysis updates, agent and long-context tests), to see whether the 43 score holds and whether the mixed results resolve.
- 02Real-world cost per completed task, with price changes from Kimi, OpenAI and Anthropic, and whether Alibaba's cache-hit and long-context pricing holds at scale.
- 03Adoption of the open-source Qwen-Live-Delivery-Plugin (GitHub activity, third-party integrations) and whether the Live Delivery feature works reliably in production agent setups.
Coverage
1 storyNewsletter
Get the no-hype AI news straight to your inbox
Weekly. No spam, unsubscribe any time. Privacy
More capability signals
Full feed →- Deepseek releases V4.1-Flash, an open-source model that sharply cuts KV cache memory and input-processing compute for AI agents10 Sept 202682
- Mistral releases Mistral Large 4, a ~1T-parameter open-weight multimodal MoE model, with day-0 Hugging Face support29 Oct 202680
- Meta releases Muse Spark Open, an open-weight ~816B-parameter multimodal mixture-of-experts model, with Hugging Face support15 Oct 202680
- OpenAI discloses sandbox-escape and credential-leak incidents, confirms pause on tool-use for its most capable models26 Sept 202680