Alibaba's Qwen3.8-Max Reaches GA With Leading 86.1 Agentic Benchmark Score

Qwen3.8-Max is a 2.4-trillion-parameter multimodal mixture-of-experts model with 95 billion active parameters developed by Alibaba's Qwen team.

Alibaba announced the general availability of Qwen3.8-Max on August 2, 2026, through its QwenCloud service on the Model Studio platform. The rollout provides immediate API access for developers working on agentic and coding applications. The model enters a competitive landscape where performance on standardized agent benchmarks determines adoption rates among enterprise users.

The release follows internal testing that positioned the model for production workloads requiring sustained interaction with desktop environments. Alibaba Cloud lists the model as a flagship offering with its specified parameter count and architecture details available in the console interface. This availability expands options for organizations already using the cloud provider's inference infrastructure.

How does Qwen3.8-Max compare on agentic benchmarks?

Performance data from the official announcement shows Qwen3.8-Max achieving 86.1 on the OSWorld-Verified agentic desktop benchmark. This metric measures the model's ability to complete verified tasks in simulated desktop settings over extended sequences. The score exceeds the 83.2 recorded by GPT-5.6 Sol Max and the 85.0 achieved by Fable 5 on the same evaluation.

The OSWorld-Verified benchmark focuses on long-horizon agentic behavior, including navigation, tool use, and multimodal input processing. Higher scores indicate stronger reliability when models must maintain context across multiple steps without human intervention. Qwen3.8-Max's result establishes a new reference point for open-weight contenders in this category.

Developers evaluating models for agent deployments can reference these figures directly when selecting between providers. The benchmark results appear in tables published alongside the general availability notice. Such comparisons help quantify the incremental gains in agent reliability that the new model delivers.

Today, we are officially releasing Qwen 3.8-Max, the most capable model in the Qwen family to date. This also marks the first time we will open-source the weights of a Qwen-Max-class model — the open weights will be released next week.QwenTeam, Alibaba Qwen team

What technical architecture supports Qwen3.8-Max?

The model employs a mixture-of-experts design that totals 2.4 trillion parameters while activating only 95 billion during each forward pass. This selective activation reduces computational overhead compared with dense models of similar total size. Multimodal support extends to text, image, and other input modalities required for agentic desktop scenarios.

The architecture prioritizes efficiency in long-context and long-horizon tasks where repeated inference calls accumulate costs. Active parameter selection occurs dynamically based on input characteristics. This mechanism allows the model to allocate resources toward relevant expert subnetworks for coding or workflow execution.

Alibaba Cloud documentation confirms the 2.4 trillion total parameter figure and the MoE structure in the Model Studio console listing. These specifications align with the model's positioning for enterprise agent applications that demand both scale and cost control.

What pricing applies to Qwen3.8-Max API usage?

Alibaba Cloud sets API rates at $2 per million input tokens and $6 per million output tokens. The tiered structure applies uniformly across input and output volumes processed through the Model Studio endpoint. Organizations can calculate expected costs based on typical token consumption patterns in agent loops.

The pricing appears in the official Model Studio pricing documentation alongside other Qwen family offerings. This rate aims to balance accessibility for smaller teams with the model's high benchmark performance. Users integrating the model into production agent systems receive predictable per-token billing.

Comparison of agentic benchmark scores and available pricing details for leading models

Model	OSWorld-Verified Score	Total Parameters	Active Parameters	Input Price per Million Tokens	Output Price per Million Tokens
Qwen3.8-Max	86.1	2.4T	95B	$2	$6
GPT-5.6 Sol Max	83.2	N/A	N/A	N/A	N/A
Fable 5	85.0	N/A	N/A	N/A	N/A

What market implications follow from the release?

The combination of benchmark leadership and accessible API pricing may accelerate adoption of Alibaba Cloud infrastructure among agent developers. Existing Qwen users gain a direct upgrade path without changing providers. The upcoming open weights release could extend influence beyond paid API customers to the broader research community.

Competitors face pressure to match or exceed the 86.1 benchmark result while maintaining similar cost structures. Enterprise stakeholders evaluating agent platforms now include Qwen3.8-Max in shortlists alongside established options. The timing of the open weights announcement may influence decisions on proprietary versus community-driven model strategies.

The model targets specific verticals including software development assistance and autonomous workflow management. These use cases benefit from the long-horizon capabilities demonstrated in the benchmark. Market analysts will track adoption metrics following the open weights availability to assess long-term competitive shifts.

What reactions have accompanied the announcement?

The Qwen team statement emphasizes the model's status as the most capable in the family and highlights the milestone of open-sourcing a Max-class model for the first time. This positions the release as both a technical and community milestone. The statement appears in the official blog post detailing the general availability.

Industry observers note the alignment between benchmark gains and the decision to release weights publicly. Such moves can accelerate downstream fine-tuning and evaluation efforts by external researchers. The pricing announcement alongside the performance claims provides concrete data points for procurement teams.

What developments are expected next for Qwen3.8-Max?

The open weights release scheduled for the week after August 2, 2026, will enable local deployment and customization. Community contributions may emerge rapidly once the weights become public. Alibaba Cloud plans to maintain API availability for users preferring managed inference.

Further updates to the Qwen series may incorporate lessons from initial production deployments of Qwen3.8-Max. The focus remains on expanding agentic and multimodal features for enterprise scenarios. Stakeholders should monitor the console and pricing pages for any adjustments following the weights release.

1. Review the official Qwen blog for open weights download instructions once available.
2. Test API endpoints in the Model Studio console to establish baseline performance.
3. Compare token consumption in specific agent workflows against current pricing.
4. Evaluate integration options with existing Alibaba Cloud services for production rollout.
5. Track subsequent benchmark updates from the Qwen team for iterative improvements.

The general availability marks the start of broader evaluation cycles across the AI development community. Performance on additional agent benchmarks beyond OSWorld-Verified will provide further context on relative strengths. The combination of technical specifications, pricing, and release timeline offers a clear path for interested parties to engage with the model.