Frontier Models

Qwen3.8-Max: Alibaba's 2.4T MoE Model Advances Autonomous Coding Frontier

The 2.4 trillion parameter model from Qwen supports extended autonomous runs and multimodal feedback loops, with immediate API access and open weights planned for the following week.

By Marcus Vance August 3, 2026 3 MIN READ

In a spacious modern open-plan technology office located within an Alibaba corporate facility the scene shows a single anonymous software engineer viewed from behind seated at a large ergonomic desk facing a bank of four flat panel computer monitors and a high performance workstation tower. The engineer wears a plain dark gray button down shirt and sits upright with hands resting on a standard keyboard and mouse pad. On the desk surface sit multiple external hard drives connected by thick black cables a small array of cooling fans with visible metal grilles and a secondary laptop closed and placed to the side. The primary workstation tower features visible internal components through a tempered glass side panel including densely packed circuit boards and heat sinks representing the computational infrastructure supporting large scale mixture of experts model training. The monitors display abstract colorful graphical interfaces with flowing lines and node connection diagrams that represent neural network architectures multimodal data streams and autonomous code generation processes without any readable characters or symbols. Behind the desk a large window reveals an adjacent server room visible through glass partitions containing rows of black server racks with blinking indicator lights and bundled fiber optic cables running along the floor and ceiling trays. The server racks symbolize the QwenCloud infrastructure hosting the 2.4 trillion parameter Qwen3.8-Max model enabling extended autonomous coding runs. On a nearby side table rest additional pieces of hardware including a stack of solid state drives and a tablet device showing similar abstract network visualizations. The overall environment includes neutral colored walls subtle overhead lighting fixtures potted plants in the distance and other empty desks with similar equipment setups creating a realistic photojournalistic view of a professional setting where engineers interact with advanced artificial intelligence tools from Qwen and Alibaba. The composition emphasizes hardware elements cables cooling systems and the solitary figure engaged in observation of the system outputs that illustrate multimodal feedback loops and frontier model capabilities in autonomous software development tasks. Additional details include the texture of the carpeted floor the arrangement of ventilation grilles in the ceiling the precise positioning of monitor stands the thickness of bundled cables snaking between equipment the reflective surfaces on the server rack doors and the subtle shadows cast across the workspace all contributing to a dense realistic live action capture of the technological environment supporting Qwen3.8-Max and related Qwen3.8-27B variants without any human faces or identifiable individuals shown. — Illustration: AI Intel Report

Qwen3.8-Max is a 2.4 trillion parameter Mixture-of-Experts model built on the Qwen 3.5 architecture and described as the most capable in the Qwen family to date.

Qwen released Qwen3.8-Max on August 2, 2026. The model is the most capable in the Qwen family to date. This marks the first time Qwen will open-source the weights of a Qwen-Max-class model.

The model demonstrates native multimodal intelligence. Vision serves as a continuous feedback loop for planning. Vision serves as a continuous feedback loop for execution. Vision serves as a continuous feedback loop for self-correction.

What technical specifications define the Qwen3.8-Max model?

Qwen3.8-Max scales to 2.4 trillion parameters. The model uses a Mixture-of-Experts architecture. The model has 95B active parameters in some references. The model supports one million token context.

The model is available via QwenCloud API. The input pricing is two dollars per million tokens. The output pricing is six dollars per million tokens.

How did Qwen3.8-Max perform in autonomous coding demonstrations?

Qwen3.8-Max performed 10 plus days of autonomous coding. The task was to build a self-evolving harness. The full project trace is available in a public GitHub repository.

Qwen3.8-Max achieved 500 plus turns of chip design optimization. The process was closed-loop. The gate count was reduced from 8,298 to 678.

Key Metrics for Qwen3.8-Max
Metric	Value	Attributed Source
Parameters	2.4 trillion	Qwen
Autonomous Coding Duration	10+ days	Qwen
Optimization Turns	500+	Qwen
Gate Count Reduction	8,298 to 678	Qwen
Input Token Price	2 USD per million	QwenCloud
Output Token Price	6 USD per million	QwenCloud

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, Qwen research team

What are the access options and planned releases for Qwen3.8-Max?

The model is available through the QwenCloud platform. Open weights for Qwen3.8-Max are scheduled for next week. Open weights for Qwen3.8-27B are also scheduled for next week.

The announcement was made on August 2, 2026.
The API is available immediately on QwenCloud.
The open weights release is planned for the following week.
The model supports native multimodal agents.
The GitHub repository provides evidence of the autonomous development.

What market implications arise from the Qwen3.8-Max release?

The release positions Qwen as a player in the frontier models space. The open weights will allow broader developer access. The pricing structure offers a specific cost for input and output tokens.

Stakeholders in AI development may examine the autonomous coding capabilities. The 10 plus day runs suggest advances in long-horizon tasks. The 500 plus turn optimizations indicate closed-loop improvement processes.

What comes next for the Qwen3.8-Max model and related releases?

The open weights will be released next week. The Qwen3.8-27B model will also have open weights. The model is built to support multimodal intelligence in agents.

Developers can access the model through the QwenCloud API in the meantime. The public GitHub repository offers ongoing commits and issues as of August 2026. The repository demonstrates autonomous development with Qwen Code.

Frequently asked

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

The open weights for Qwen3.8-Max are scheduled for release next week following the August 2, 2026 announcement.

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

The pricing is two dollars per million input tokens and six dollars per million output tokens on QwenCloud.

Sources

Qwen — Qwen3.8-Max is a 2.4 trillion parameter model with 10+ day autonomous coding and 500 turns optimization.
QwenCloud — The Qwen3.8-Max API pricing is two dollars per million input tokens and six dollars per million output tokens.
GitHub — The GitHub repository shows a minimal autonomous code-agent CLI built with Qwen Code with ongoing commits and issues as of August 2026.
X — 📢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!