Qwen3.8-27B: Qwen's Dense Vision-Language Model With a 262K-Token Native Context

Qwen has published details on Qwen3.8-27B, a 27-billion-parameter dense causal language model paired with a vision encoder, supporting both image and video understanding alongside a native context length of 262,144 tokens that can be extended to 1,000,000.

• 27-billion-parameter dense model with an integrated vision encoder for image and video understanding.
• Native context length of 262,144 tokens, with support for extension to 1,000,000 tokens.
• Thinking mode is enabled by default, with reasoning depth adjustable via a reasoning_effort parameter.
• Preserved thinking is enabled by default alongside the thinking mode.
• Architecture combines gated DeltaNet linear attention heads with gated attention heads across 64 layers and a 5120 hidden dimension.
• Confirmed support for Hugging Face Transformers, vLLM, SGLang, and TokenSpeed.

• Developers building applications that need a single model to reason over text, images, and video.
• Teams evaluating dense architectures with configurable reasoning depth rather than fixed chain-of-thought behavior.
• Engineers already working with Hugging Face Transformers, vLLM, or SGLang who want to test a model that plugs into those existing serving stacks.
• Researchers interested in gated DeltaNet linear attention combined with gated attention mechanisms within one architecture.

Qwen's blog post at qwen.ai/blog?id=qwen3.8 lists supported inference frameworks — Hugging Face Transformers, vLLM, SGLang, and TokenSpeed — as the confirmed access paths for running Qwen3.8-27B. Consult that post directly for the exact model identifier, checkpoint location, and framework-specific loading instructions before attempting a deployment, since those details are not specified beyond the framework names themselves.

Sources

• Qwen3.8 announcement — Qwen