On August 12, 2026, Alibaba's Qwen team released Qwen3.8-2.4T-A95B, a mixture-of-experts (MoE) model with 2.4 trillion total parameters and 95 billion active per token, and followed it two days later with the dense Qwen3.8-27B. The team says Qwen3.8 brings a "Qwen-Max-class" model to open release for the first time, a notable step because the Max tier had previously been reserved for Alibaba's hosted API. For developers, Qwen3.8 adds a frontier-scale option to the most widely used open model family, but also a license that needs closer reading than earlier Qwen releases.

Key Facts#

  • Release dates: Qwen3.8-2.4T-A95B on August 12, 2026, and Qwen3.8-27B on August 14, 2026, both on Hugging Face and ModelScope.
  • Positioning: the Qwen team says that "for the first time, Qwen3.8 brings a Qwen-Max-class model to open release."
  • Foundation: built on the Qwen3.5 architecture, which combines Gated Delta Networks (a linear-attention design) with sparse MoE layers and unified vision-language training.
  • Focus: coding, professional work, research and long-horizon agentic tasks, with better planning and handling of environment feedback, according to the team.
  • Reasoning controls: reasoning depth is set with reasoning_effort, and preserve_thinking keeps reasoning from earlier messages in context.
  • Serving: documented for vLLM, SGLang, TokenSpeed and Transformers, with OpenAI-compatible endpoints and examples using a 262,144-token context.
  • License: distributed with each checkpoint rather than as a blanket Apache 2.0 grant. Hugging Face reported that the 2.4T model's license adds some non-commercial restrictions and revenue-share requirements.

What Happened#

Qwen3.8 is the third major open release in the Qwen3.5 series in 2026. The Qwen team describes the goal as carrying complex, multi-step tasks through to completion more reliably, not only answering harder questions. Its README lists four areas of improvement: core capabilities across coding and professional work, stronger autonomous planning and use of environment feedback, broader compatibility with popular agent harnesses and developer tools, and more flexible thinking control.

The deployment guidance shows how the team expects developers to use the models. Each server recipe enables a Qwen reasoning parser and a qwen3_coder tool-call parser, and exposes an OpenAI-compatible API. The team also highlights hosted access through its QwenCloud API, which it says is compatible with both OpenAI and Anthropic API specifications, and through its own Qwen Code terminal agent.

The team did not publish benchmark tables in the repository for Qwen3.8. It points to the model cards on Hugging Face for detailed results, which means independent evaluations will matter more than usual for this release.

Background#

The Qwen3.5 generation began on February 16, 2026, with a 397B-A17B MoE model, followed by 122B, 35B and 27B models later that month and small models down to 0.8B in early March. Qwen3.5 introduced a unified vision-language foundation trained with early fusion, support for 201 languages and dialects, and a hybrid architecture that pairs Gated Delta Networks with sparse MoE for faster inference. CNBC framed the launch as part of China's chatbot race shifting toward AI agents. Qwen3.6 followed in April 2026 with 35B-A3B and 27B models focused on agentic coding and on preserving reasoning across conversation history.

By the summer, Qwen had become the community's default base model. Hugging Face's summer 2026 report counted more than 151,000 Qwen-derived models on the Hub, 2.6 times Meta's entire footprint, and about 39.6 million Qwen GGUF downloads a month. Even the README of OllamaSharp, a popular .NET client for Ollama, uses a Qwen3.5 model in its getting-started example. That context explains why the license on Qwen3.8's largest model drew attention. Earlier Qwen generations, including the original Qwen3 release in April 2025, shipped all open weights under Apache 2.0.

Why It Matters for Developers#

The 27B model is the practical story for most teams. A dense 27B model with Qwen3.8's agentic tuning fits on a single high-memory GPU with quantization or shorter contexts, while the documented recipes use four GPUs for the full 262K context. Those vLLM and SGLang commands give you an OpenAI-compatible endpoint that .NET code can reach through Microsoft.Extensions.AI or the OpenAI library. Use the recommended reasoning and tool-call parsers. If you serve the model without them, reasoning text and tool calls can leak into the visible response, and agent frameworks will misinterpret the output.

Reasoning retention is becoming a standard feature. Like Moonshot's Kimi K3, Qwen3.8 can keep reasoning from earlier turns through preserve_thinking. That improves multi-step agents, but only if your client and chat-history storage preserve the reasoning content. Verify this in integration tests before relying on it, and set reasoning_effort per scenario to keep latency and token costs predictable. Our guide to function calling in .NET shows how to build that kind of regression suite.

Review licenses per checkpoint. With Qwen3 you could assume Apache 2.0 across the family. With Qwen3.8, the terms travel with each model, and the flagship's reported conditions could affect commercial platforms that resell model access. Record the exact license file you accepted as part of your model inventory, the same way you track NuGet package licenses.

The 2.4T model itself is mainly relevant through APIs and certified hosts. Self-hosting a 2.4T-parameter MoE model takes cluster-scale hardware, while the hosted QwenCloud endpoint offers the same model family through familiar API shapes.

What's Next#

Qwen's cadence suggests more releases soon, and the Qwen organization on GitHub has already added a Qwen3.8-Flash-Next repository for a new foundation model. The open questions are how Qwen3.8 performs in independent evaluations, whether the smaller Qwen3.8 models keep permissive terms, and whether other labs follow Moonshot and Alibaba in attaching commercial conditions to their largest open weights.

For .NET teams, the strategy is the same as with the original Qwen3 release: keep model access behind an abstraction, benchmark candidates on your own tasks, and treat license terms as part of model selection. Our guide to local AI in .NET covers the runtimes that make switching easy, and our agent architecture guide covers how to design for long-running, multi-step tasks.

Sources#