Qwen Launches Qwen3.8 27B, an Open-Weight Vision-Language Model with 262K Context

Qwen has released Qwen3.8 27B, a new open-weight dense vision-language model, now accessible through OpenRouter's API as qwen/qwen3.8-27b.

Key specs

• Context window: 262,144 tokens (262K)
• Pricing: $0.45 per 1M input tokens, $3.20 per 1M output tokens
• Modality: text + image + video → text
• Architecture: dense (non-mixture-of-experts) vision-language model
• Parameters: 27 billion

What the model does

According to Qwen, Qwen3.8 27B is built for coding, professional workflows, research tasks, multimodal interaction, and long-running agent tasks. The company describes the model as having "flexible thinking" — language suggesting a configurable reasoning or chain-of-thought mode, though Qwen has not published details on how this mode is toggled or how it affects latency and output length.

The 262K context window places Qwen3.8 27B among the longer-context open-weight models currently available, well beyond the 128K windows common in many competing mid-size models. Combined with multimodal input support for images and video, the model is positioned for tasks that require ingesting large documents, codebases, or video content alongside visual reasoning.

No benchmark scores — MMLU, HumanEval, or otherwise — have been disclosed alongside this release. Qwen has not published a technical report or model card detailing training data, training cutoff date, or evaluation results at time of writing.

Availability

The model is live now on OpenRouter, priced at $0.45 per million input tokens and $3.20 per million output tokens. As an open-weight release, it is expected to also become available for self-hosting or through additional inference providers, following the pattern of prior Qwen model launches, though Qwen has not confirmed weight availability details in the source material reviewed for this article.

What this means

Qwen3.8 27B extends the company's open-weight lineup into a dense vision-language configuration with an unusually large context window for its parameter class. The 27B size sits in a competitive middle tier — large enough for meaningful multimodal reasoning, small enough to run on a single high-memory GPU or modest multi-GPU setup, which matters for teams that want to self-host rather than rely on API access.

The lack of published benchmarks is a real gap. Buyers evaluating this model for coding or agentic workloads will need to run their own evals against comparable open models — such as other dense VLMs in the 20B-30B range — before committing production traffic. The pricing, at $0.45/$3.20 per million tokens, undercuts many closed-source multimodal APIs, which is consistent with Qwen's broader strategy of using aggressive open-weight pricing to build developer adoption ahead of proven benchmark parity. Until independent evaluations surface, treat the "suited for professional workflows and long-running agent tasks" description as a company claim rather than a verified capability.