EDGE OS
Alibaba’s Qwen team released Qwen3.8-27B on August 14, 2026, and it quickly became one of the most downloaded open-weight models on Hugging Face. It is a dense vision-language model that is small enough to run on a single GPU, which makes it a practical choice for companies that want a capable private AI model on their own hardware.
This guide covers how big Qwen 3.8 27B is, what it can do, its license and context window, and the hardware you need to run it locally.
What is Qwen 3.8 27B?
Qwen3.8-27B is the smaller companion to Qwen3.8-Max, released two days after Max’s weights. Unlike most large open models today, it is dense rather than a mixture of experts: every parameter works on every token. It reads text, images and video, and writes text.
- Parameters: about 27.8 billion, including the vision encoder
- Architecture: 64 layers, hidden size 5,120, with three linear-attention layers for every full-attention layer
- Input and output: text, image and video in; text out
- Context window: 262,144 tokens natively, extendable to 1 million
- License: Apache 2.0, commercial use allowed
- Released formats: BF16, FP8 and an NVFP4 build of 20.4 GB
How big is Qwen 3.8 27B?
Size is the most searched question about this model. In memory terms, what matters is how many bytes each parameter takes, which depends on the precision you run it at:
| Precision | Weights | 1 user, 8K context | 16 users, 8K context |
|---|---|---|---|
| BF16 | 56 GB | 63 GB | 73 GB |
| FP8 | 28 GB | 32 GB | 43 GB |
| NVFP4 (4-bit) | 21 GB | 24 GB | 35 GB |
Each user also needs their own KV cache, which grows with context length. Qwen 3.8 27B keeps this small because only a quarter of its layers use full attention: about 64 KB per token, so an 8K conversation adds roughly half a gigabyte.
What can Qwen 3.8 27B do?
Alibaba positions Qwen3.8-27B for coding, agents and computer use. In Alibaba’s own evaluations, run with a 256K context, it scored:
- 61.7 on SWE-bench Pro (software engineering)
- 84.3 on OSWorld-Verified (computer use)
- 89.2 on GPQA Diamond (graduate-level science questions)
- 90.0 on MathVision (visual math)
These are vendor-reported results. Alibaba adjusted some benchmark tasks, so the scores are not directly comparable with unmodified versions. Test it on your own documents and prompts before deciding.
Is Qwen 3.8 27B free and open source?
Yes. The weights are published on Hugging Face and ModelScope under the Apache 2.0 license, which allows commercial use without a revenue share. You pay nothing for the model itself; your costs are the hardware and power to run it.
How to run Qwen 3.8 27B locally
Because it fits on one GPU, Qwen 3.8 27B is one of the easiest strong models to self-host. To run it on your own hardware you need:
- GPU memory: about 32 GB at FP8 or 24 GB at NVFP4 for one user with 8K context, and about 43 GB for a team of 16 at FP8
- System RAM: at least as much as the model files, about 28 GB at FP8
- Software: an inference engine such as vLLM or SGLang, both of which publish serving recipes for Qwen3.8-27B
- The model: the weights from Hugging Face (Qwen/Qwen3.8-27B or Qwen/Qwen3.8-27B-FP8)
For a team of 16 at FP8, each of these single GPUs is enough: NVIDIA L40S, NVIDIA H100 NVL, NVIDIA RTX PRO 6000 Blackwell, NVIDIA H200 NVL, AMD Instinct MI300X, AMD Instinct MI325X. The NVIDIA DGX Spark desktop system also fits.
Our Qwen 3.8 27B hardware requirements page has VRAM for every precision, user count and context length, plus fine-tuning needs. Or adjust the numbers yourself in the LLM VRAM calculator.
Running Qwen 3.8 27B privately with Vanik Edge
Vanik Edge OS runs open-weight models like Qwen 3.8 27B on GPU servers inside your building. Your team gets a private ChatGPT-style assistant and an OpenAI-compatible API, and documents, prompts and outputs never leave your network.
Frequently asked questions
How big is Qwen 3.8 27B?
Qwen 3.8 27B has about 27.8 billion parameters including its vision encoder. The weights take about 56 GB in BF16, 28 GB in FP8 and 20.4 GB in the official NVFP4 build.
Is Qwen 3.8 27B out?
Yes. Alibaba released Qwen3.8-27B on August 14, 2026, with weights on Hugging Face and ModelScope.
Is Qwen 3.8 free and open source?
Yes. Qwen3.8-27B is released under the Apache 2.0 license, so you can download it, self-host it and use it commercially at no cost.
Does Qwen 3.8 have vision?
Yes. Qwen3.8-27B is a native vision-language model that takes text, image and video input and produces text.
What's the max context window for Qwen 3.8?
Qwen3.8-27B supports 262,144 tokens natively. Alibaba says it can extend to 1 million tokens with YaRN, though static YaRN in common open-source frameworks can reduce accuracy on shorter inputs.
Can I run Qwen 3.8 27B locally?
Yes. At FP8 it needs about 32 GB of GPU memory for one user with 8K context, and about 24 GB at NVFP4, so a single 48 GB data center GPU runs it comfortably. Serving a team of 16 at FP8 needs about 43 GB.
What is the best computer for Qwen 3.8 27B?
For one person, any machine with a GPU of 32 GB or more works at FP8. For a team, a server with an NVIDIA L40S, RTX PRO 6000 Blackwell or H100-class GPU gives room for many users and longer context.
Run open-weight models inside your building
Tell us the models and team size you have in mind. We will size the GPU server with you and set up Vanik Edge OS on it.