Qwen3.8-27B goes open: a multimodal 27B model with a roughly 50GB local footprint (2026)

Qwen's promised open-weight follow-up is now a download rather than a release countdown. The company's verified account announced Qwen3.8-27B at 15:02 on August 14, linking both Hugging Face and ModelScope. 1

Field	What is confirmed
Release	Qwen3.8-27B is a 27B dense causal model with a vision encoder, released under Apache 2.0. 2 3
Local footprint	The Hugging Face repository contains 18 BF16 `safetensors` shards totaling roughly 50GB. The model card describes image and video understanding, coding, professional work, research, and long-horizon agent tasks. 2
Context and reasoning	Native context is 262,144 tokens. Qwen says it can be extended toward 1M with RoPE/YaRN scaling; thinking is enabled by default and exposes `xhigh`, `medium`, and `low` reasoning effort. 2
Action window	Start a local evaluation now if your stack can handle the memory; Qwen Cloud's hosted version is documented as coming soon. 2

What changed

The important change is deployment scope. Qwen3.8-27B is a dense, downloadable model rather than the much larger hosted Qwen3.8-Max line. Its vision encoder accepts images and videos, while the reasoning controls let developers trade latency and answer depth per request. The repository documents compatibility with Transformers, vLLM, SGLang, and TokenSpeed, so the first path to production is assembling an inference stack around the weights rather than waiting for a managed endpoint. 2

What the numbers say

Qwen reports 73.0 versus 63.4 for Qwen3.6-27B on Terminal Bench 2.1, 61.7 versus 53.5 on SWE-bench Pro, and 42.2 versus 13.3 on DeepSWE 1.1. Those are useful signals for coding and agent work, but they are launch evidence from Qwen: the card says SWE-bench Pro and DeepSWE used the Claude Code harness, while QwenSWEBench and CoWorkBench are in-house evaluations. They should guide what to test, not settle the ranking. 2

The constraint is memory

The official BF16 repository is far heavier than the 27B name suggests: a roughly 50GB download is before runtime overhead, KV cache, or long-context costs. LM Studio lists a 17GB quantized package and minimum RAM guidance for its own local distribution, while AMD says its supported path needs roughly 24GB of VGM or VRAM to run comfortably. AMD's preliminary tests reached up to 24.5 tokens per second on a Ryzen AI Max+ 395 and 51.8 on a Radeon AI PRO R9700, but those are hardware-vendor measurements, not model-quality results. 4 5

For developers, the release changes the next step from watching to testing: put the weights behind a real tool loop, repository, or document workflow, then measure latency, memory use, and task completion against the model you already run. The breadth of Qwen's table makes the release worth an evaluation; the hardware bill and still-pending independent validation decide whether it stays in your stack.