Frontier Models

Alibaba Launches Qwen3.8-Max 2.4T MoE Model with Open Weights Next Week

The release positions a 2.4 trillion parameter sparse MoE model on QwenCloud at $2 input and $6 output per million tokens while scheduling open weights for the Max variant and the 27B model within one week of the August 2, 2026 announcement.

By Marcus Vance August 3, 2026 4 MIN READ

Inside a vast secure Alibaba data center facility dedicated to QwenCloud operations multiple rows of tall black server racks stretch into the distance each rack filled with densely packed GPU accelerator modules and networking switches illuminated by rows of small green status LEDs indicating active computation clusters supporting large scale sparse mixture of experts inference workloads generic anonymous technicians wearing white lab coats ear protection and ID badges with backs turned to the viewer stand at open rack doors using tablet devices to monitor system health metrics one technician adjusts cabling connections on a lower shelf while another points toward a diagnostic display on an adjacent rack wall mounted cooling units hum with visible condensation lines running along overhead pipes and industrial grade power distribution units line the concrete floor with heavy duty cables neatly bundled and routed through raised floor grates reflective polished concrete surfaces show subtle reflections of the rack lights creating depth in the industrial environment air filtration vents are visible high on the walls alongside emergency signage symbols without any lettering large metal support columns divide the space and safety barriers cordon off restricted zones in the background a separate area holds test benches with disassembled hardware components including cooling plates and interface boards arranged for inspection the overall atmosphere conveys high precision engineering activity centered on deploying and maintaining the Qwen3.8-Max 2.4 trillion parameter model alongside the smaller Qwen3.8-27B variant both scheduled for open weight release the scene includes subtle branding elements such as stylized cloud motifs etched into rack panels representing QwenCloud infrastructure without any readable text or logos multiple layers of cabling trays overhead carry fiber optic lines and power conduits contributing to the dense technical environment generic figures remain faceless and non identifiable emphasizing the hardware and operational setting of frontier model infrastructure rollout on August timelines the composition focuses on authentic industrial details like torque wrenches on tool carts spare component bins labeled with color codes only and environmental sensors mounted on pillars capturing the real world scale of cloud based AI model hosting and preparation for public weight availability — Illustration: AI Intel Report

Qwen3.8-Max is a 2.4 trillion parameter sparse mixture-of-experts model developed by Alibaba featuring approximately 95 billion active parameters.

The model was officially released on August 2, 2026, as the most capable in the Qwen family to date.

This release marks the first instance of open-sourcing weights from a Qwen-Max class model.

What background led to the Qwen3.8-Max release?

Alibaba has developed the Qwen family of models through successive iterations focused on coding and collaborative tasks.

The progression within the family has emphasized increasing model scale and capability over time.

The current model builds directly on that lineage with an expanded architecture and new distribution plans.

The announcement includes a specific timeline for open weights availability that differs from prior releases in the series.

What are the technical specifications of Qwen3.8-Max?

Qwen3.8-Max is a 2.4 trillion parameter sparse MoE model with approximately 95 billion active parameters.

The sparse mixture of experts design supports the total parameter count while limiting active parameters during use.

The model demonstrates long-horizon autonomous capabilities including 10+ day self-evolving coding projects and multi-day research reproduction with self-improvement.

These capabilities extend the range of tasks the model can perform without continuous human intervention.

How is Qwen3.8-Max made available to users?

The model is now available via QwenCloud for immediate API access.

Input pricing stands at $2.0 per million tokens and output pricing stands at $6.0 per million tokens.

Implicit caching carries a separate rate of $0.25 per million tokens.

This structure provides options for different usage patterns and cost considerations.

What benchmarks does the model achieve?

Qwen3.8-Max benchmark performance
Benchmark	Score
TerminalBench 2.1	86.6
PaperBench	93.0
FrontierSWE	73.5
CoWorkBench	74.8

The TerminalBench 2.1 score of 86.6 reflects performance on terminal based evaluation tasks.

The PaperBench score of 93.0 indicates results on paper related benchmarks.

FrontierSWE at 73.5 and CoWorkBench at 74.8 provide additional measures of software engineering and collaboration performance.

What is the timeline for the open weights release?

Official release occurred on August 2, 2026.
API access is currently available on QwenCloud.
Open weights for Qwen3.8-Max and Qwen3.8-27B expected next week.

The ordered sequence of events begins with the initial announcement and API availability.

The subsequent step involves the public distribution of model weights for both variants.

What market implications follow from this release?

The pricing levels may affect how users compare offerings across different providers.

The open weights commitment for a Max class model introduces a new option for local deployment.

Stakeholders such as developers and researchers gain expanded choices for model access.

Enterprise users may evaluate the API and open weights paths based on their infrastructure needs.

The long horizon capabilities point to potential uses in extended autonomous operations across industries.

What expert reactions have been noted regarding Qwen3.8-Max?

Meet Qwen3.8-Max — our most capable model to date. Next week, the open weights of Qwen3.8-Max will be released, and Qwen3.8-27B is also going open-weights to meet you all!Qwen team

The Qwen team statement emphasizes the model as the most capable to date in the family.

The announcement links the capability claim directly to the parameter scale and open weights plan.

The focus on coding and cowork benchmarks aligns with the stated use case priorities.

What developments are expected next for the Qwen series?

The open weights release next week will enable community access and potential fine tuning efforts.

The Qwen3.8-27B variant will accompany the Max model in the open weights distribution.

Users may explore applications that leverage the demonstrated self evolving and research reproduction features.

Subsequent updates to the Qwen family could build on the architecture introduced in this release.

The dual availability of API and open weights provides flexibility for varied deployment scenarios.

Frequently asked

When will the open weights of Qwen3.8-Max be released?

The open weights of Qwen3.8-Max and Qwen3.8-27B are scheduled for release next week after the official launch on August 2, 2026.

What is the API pricing for Qwen3.8-Max?

Input pricing is set at $2.0 per million tokens and output pricing is set at $6.0 per million tokens with implicit caching at $0.25 per million tokens.

Sources

Qwen — Announcement of the Qwen3.8-Max release as the most capable model and the parameter count of 2.4 trillion.
Alibaba_Qwen — Pricing details and confirmation of open weights release for Qwen3.8-Max and Qwen3.8-27B.