What was released
Alibaba's Qwen team has released open weights for its Qwen3.8 model family under the Apache 2.0 license, with the weights published on Hugging Face and ModelScope. The core release is the Qwen3.8-27B, a dense multimodal model with 27 billion parameters, alongside a much larger Max-level model designated Qwen3.8-2.4T-A95B, which the team built to operate at the Max level.
What the 27B model does
According to Qwen, the Qwen3.8-27B outperforms the larger Qwen3.7-Plus on coding and office tasks and is positioned for local inference on consumer hardware, targeting setups with around 24GB of VRAM such as an NVIDIA RTX 4090. The model natively processes text, images, and video — including diagrams, documents, and multi-hour video — and supports use cases such as automated video analysis, visual question answering, document understanding with charts, and multimodal coding assistants that interpret screenshots. It also offers improved agent behavior, with more independent planning and more reliable task completion.
Context length and operating modes
The Qwen3.8-27B natively handles up to 262,000 tokens of context and can be scaled to one million tokens using the YaRN method. A flexible thinking mode is enabled by default but can be toggled per query. A hosted version with one million tokens of context is planned for release through Qwen Cloud, Alibaba's AI service.
The larger Max-level sibling
Beyond the 27B model, the team released weights for Qwen3.8-2.4T-A95B, which features 2.4 trillion total parameters with roughly 95 billion active at any given time. That variant is aimed at enterprise-scale, multi-node data center deployments rather than local use, and its open-weight release extends the same Apache 2.0 terms to the top of the lineup.
Strategic shift toward open weights
Before the Qwen3.8 announcement, several Max-class Qwen models had been kept under proprietary access. Publishing both the 27B and the Max-level weights under Apache 2.0 marks a return to the open-sourcing approach that characterized earlier Qwen releases. Local deployment removes cloud dependency, reduces latency, keeps sensitive data on-premise, and lowers experimentation costs — factors that are particularly relevant for startups in regulated industries such as healthcare and finance where on-premise inference is often a compliance requirement.
What remains uncertain
Qwen3.8-27B was first announced on August 3, 2026, and by August 14, 2026, the Max-level weights had become available; coverage notes the smaller 27B model's formal release listing had not yet appeared at the time of reporting, with community sources pointing to a scheduled launch. The exact timing of the Qwen Cloud hosted one-million-token rollout was not specified beyond a "soon" indication from the Qwen team.
Share this article
