# Sam Witteveen — 'Qwen3.8-27B & How to Serve it Fast'

> A walkthrough of Qwen3.8-27B plus the serving stacks that keep the open-weights vision model quick.

- Org: Sam Witteveen
- Date: 2026-08-18
- Category: video
- Source: https://www.youtube.com/watch?v=PTuGGdDuyPI
- Canonical page: https://ai-tldr.dev/releases/sam-witteveen-qwen-3-8-27b-serving-aug18/

## In short

Sam Witteveen walks through Qwen3.8-27B, Qwen's Apache-2.0 vision-language model, and how to serve it fast. The dense 27B model handles 262,144 tokens of context natively and stretches to about 1M.

## What it is

Sam Witteveen's new video pairs a look at Qwen3.8-27B with the practical question its title asks: how to serve it fast. Qwen3.8-27B is a dense 27B vision-language model released under Apache 2.0, so teams can self-host it and read images and video alongside text without a licence fee.

## How it works

Under the hood, Qwen3.8-27B combines Gated DeltaNet with Gated Attention across 64 layers and a 5,120 hidden dimension, and is trained with multi-token prediction. Native context runs to 262,144 tokens and extends to roughly 1M. Qwen's own model card points serving at SGLang, vLLM, or TokenSpeed, and warns that throughput varies a lot between frameworks.

## Why it matters

Which server you pick decides whether a 27B open model is cheap or painful to run, so the framework half of Sam Witteveen's video matters as much as the benchmark half. Anyone weighing Qwen3.8-27B against a hosted API gets a concrete starting point for the self-hosting side of that comparison.

## Who it's for

engineers self-hosting open models

## Try it

https://huggingface.co/Qwen/Qwen3.8-27B

## Links

- [Watch on YouTube](https://www.youtube.com/watch?v=PTuGGdDuyPI)
- [Model Card](https://huggingface.co/Qwen/Qwen3.8-27B)
- [Channel](https://www.youtube.com/@samwitteveenai)