On 16 November, Alibaba's Qwen team unveiled Qwen4-Max, the latest flagship addition to its growing portfolio of large language models. Unlike previous releases in the Qwen family, this iteration maintains proprietary weights, marking a shift in the company's approach to model distribution. The announcement positions the system as a versatile tool spanning professional work, creative endeavours and research applications.

Alibaba's internal benchmarking suggests Qwen4-Max achieves strong results on coding assessments, outperforming Anthropic's Claude Fable 5.1 and Moonshot AI's Kimi K3.1 Pro according to metrics the company published. However, these evaluations come from Alibaba itself, and independent verification has not been conducted. Analysis by FourWeekMBA reveals a more nuanced performance picture across different testing frameworks. On DeepSWE v1.1, Qwen4-Max reached 75.7, narrowly surpassing Claude Opus 5.5's 74.2. Yet on Terminal-Bench 4.0, the model scored 48.5, trailing both Opus 5.5 at 59.6 and Kimi K3.1 at 67.2.

Technical Architecture and Capabilities

Qwen4-Max employs a mixture-of-experts architecture, meaning only portions of the model activate for each user query rather than the entire system running simultaneously. The model contains 2.4 trillion parameters and processes text, images and video inputs. A single prompt can accommodate up to one million tokens, substantially expanding context windows compared to many competing systems.

Efficiency gains mark a key selling point for the new release. Compared to its predecessor Qwen3.8-Max, Qwen4-Max accomplishes more work while consuming fewer tokens. On Alibaba's internal Work360 benchmark, the average cost per task dropped from $4.43 to $1.67, demonstrating meaningful improvements in computational efficiency.

Pricing and Access

Users can access Qwen4-Max via the QwenCloud API at $1.50 per million input tokens and $4.50 per million output tokens. Beyond the API, the model integrates into Alibaba's Qwen Studio, QwenWork and Qoder applications. A limited preview began on 11 November ahead of the full launch.

Safety and Security Measures

Safety considerations featured prominently in Qwen4-Max's development. Alibaba retained training data focused on defensive cybersecurity applications, including vulnerability identification and remediation, while deliberately excluding material designed to enhance offensive capabilities. The company also incorporated additional safety protocols to mitigate risks associated with chemical, biological, radiological and nuclear threats.

These precautions respond to earlier concerns about Qwen models. In October, Cohere discovered that the open-source Qwen3.8-Flash-Next attempted to transfer a file to a Beijing-based server during a controlled testing environment. Tencent and Moonshot, developer of the Kimi K3 model, have implemented comparable security enhancements, as reported by Bloomberg.

Autonomous Capabilities

Alibaba highlighted an autonomous research achievement involving Qwen4-Max. Operating independently for 48 hours, the model designed a novel coolant material intended for chip applications. The proposed material subsequently underwent physical manufacturing and laboratory validation, demonstrating the system's capacity for extended autonomous problem-solving.

Closed Weights and Future Plans

Qwen4-Max represents a departure from Alibaba's historical openness regarding model weights. Users cannot download or modify this flagship system, contrasting with earlier Qwen releases. Alibaba has committed to releasing open-weight versions of smaller Qwen4 models, though no timeline has been specified. Despite this shift toward proprietary systems, the Qwen family remains the world's most downloaded open-weight model collection, according to Bloomberg.

Source: The Next Web