Qwen3.8 is the latest generation of Alibaba's Qwen open models, and its 27-billion-parameter version is aimed squarely at people running AI on their own hardware. Qwen describes Qwen3.8-27B as a compact, "deployment-friendly" dense model that understands images and video as well as text, with adjustable thinking and improvements in coding and multi-step agent tasks.
Forge PC has not tested Qwen3.8 on specific hardware, and the capability claims below are Qwen's own. This guide is based on the Qwen3.8-27B model card, the Ollama and LM Studio model pages and Ollama's documentation, checked on 15 November 2026.
What Qwen3.8-27B is
Dense, not mixture-of-experts. All 27 billion parameters are used for every token. Our MoE vs dense explainer explains why that matters for speed and memory.
Vision built in. It accepts text, images and video.
Thinking on by default. Qwen says thinking mode is enabled by default but can be turned off per request, and reasoning depth can be tuned.
Long context. The model card lists 262,144 tokens natively, extendable to about one million.
Open licence. LM Studio lists it under the Apache 2.0 licence.
Tool use. LM Studio lists tool use, vision and reasoning support, so it can also use tools such as MCP servers in LM Studio.
How much memory do you need?
The size of the download is the starting point:
Ollama tag | Download size | Notes |
|---|
qwen3.8 (same as qwen3.8:27b) | 18GB | Default 4-bit build |
qwen3.8:27b-q8_0 | 30GB | 8-bit, higher quality, much bigger |
qwen3.8:27b-bf16 | 56GB | Full precision |
27b-mlx variants | 18GB and up | For Apple silicon Macs |
Sizes as listed on Ollama's Qwen3.8 tags page.
The context window (your conversation and any documents or tool results) needs memory on top of the model itself, so in practice:
24GB or 32GB graphics card: the 18GB build has room to load fully on the GPU with a moderate context. This is the comfortable case, and why 24GB cards are popular for local AI (see our best GPU for local LLMs).
16GB graphics card: the 18GB build won't fit entirely in VRAM. Ollama will split it between the GPU and system memory, which works but is slower. Make sure you have plenty of system RAM; our guide to how much RAM you need for local AI explains why 32GB is a sensible minimum and 64GB is better for models that spill over. A kit such as the Kingston FURY Beast 64GB DDR5-6000 CL30 or Crucial's Pro 64GB DDR5-6000 kit is the kind of upgrade that helps here.
8GB to 12GB graphics card: this is not the model for you. Our guide to local LLMs on 8GB of VRAM covers smaller models that fit.
Model files add up quickly, too. The three main Ollama builds alone come to over 100GB, so a large, fast SSD such as a 4TB drive (Crucial's P310 4TB, for example) makes sense if you collect models.
Option 1: Run it in Ollama
Install or update Ollama from ollama.com, as covered in our Ollama and LM Studio guide.
Open a terminal and run:
ollama run qwen3.8
Ollama downloads the 18GB default build the first time, then opens a chat. You can paste an image path into the prompt to use its vision features, as explained in our guide to vision models in Ollama.
In a second terminal, run ollama ps to see where the model loaded. Ollama's documentation explains that "100% GPU" means it's entirely in graphics memory, while a split such as "48%/52% CPU/GPU" means part of it is in system memory.
Set the context length deliberately. Ollama's documentation says it uses a 4,096-token context by default. To change it for a session, type /set parameter num_ctx 16384 in the chat (pick the size you need), or set the OLLAMA_CONTEXT_LENGTH environment variable before starting the server. Bigger contexts use more memory, so raise it in steps and recheck ollama ps.
Thinking controls. Ollama's documentation says thinking controls vary by model, and that apps can check which values a model supports through the /api/show endpoint and set think in API requests. If you only want quick answers, turning thinking off saves time on simple questions.
Option 2: Run it in LM Studio
Open LM Studio, search for Qwen3.8 and pick a download. LM Studio says it's available in GGUF and MLX formats; on a Windows PC, choose a GGUF build.
LM Studio's documentation suggests choosing a 4-bit option or higher if your machine can run it. On a 16GB card, a lower-bit file may fit fully on the GPU where a 4-bit one doesn't, at some cost in quality; our GGUF quantisation explainer covers the trade-off.
Load the model, set the context length in the load settings, and start chatting. To use it from other apps, see our LM Studio server guide.
Tips for better results
Close VRAM-hungry apps such as games and browsers with hardware acceleration before loading the model; our guide to checking VRAM usage shows what is using memory.
Treat benchmark claims with care. Qwen publishes strong scores for coding and agent tasks, but they are Qwen's own results using specific test setups.
Speed it up with the steps in our local LLM speed guide.
Our verdict
Qwen3.8-27B is a strong choice for local AI if you have a 24GB or larger graphics card: run ollama run qwen3.8 or download a GGUF build in LM Studio, and the 18GB default build should load fully with a moderate context. On a 16GB card it still runs but spills into system RAM, so make sure you have 32GB to 64GB of memory or pick a smaller quantisation. Set the context length on purpose and check ollama ps to see where it loaded. Forge PC has not tested Qwen3.8; details are from Qwen, Ollama and LM Studio.