Running Qwen3.8 locally does not have to mean renting a cloud GPU or paying per token. The open-weights Qwen3.8-27B model is free to download under the Apache 2.0 licence, and free tools such as Ollama and LM Studio can run it on hardware you already own. The real cost is download size, disk space, and patience while you choose a version that fits your memory. This guide shows the free-only route, from hardware checks to a working local chat and API.
What free actually means here
Free means no model licence fee, no API subscription, and no paid inference service. It does not mean zero resources. You still need a computer with enough memory, roughly 18 GB of download space for the common 4-bit version, and electricity while the model runs. You also need an internet connection for the one-time download; after that, chat can work fully offline.
Be careful about one common confusion. Qwen3.8-Max is a much larger cloud model in the same family, while Qwen3.8-27B is the open-weights model designed for local use. If a page asks you to buy API credits, it is almost certainly talking about a hosted model, not the local 27B download. For this guide, stick to Qwen3.8-27B and the free runtimes below.
Step 1: Check whether your hardware fits
Before downloading anything, check your memory. The full-precision model is far too large for a normal desktop, so local users run a quantized version, which compresses the weights at some cost to quality.
As a practical guide, the 4-bit Q4_K_M build downloads at about 18 GB and needs roughly 21 to 22 GB of total memory with a 32K context. That makes a 24 GB graphics card, such as an RTX 3090 or RTX 4090, a comfortable fit. A Mac with 32 GB of unified memory can also work, especially with a shorter context, while a Mac with 48 GB or more has more headroom.
If you only have a 16 GB graphics card, do not start with the 4-bit build. A smaller 3-bit quant can fit in roughly 13 to 15 GB of GPU memory, though answers may be slower and slightly less accurate. A card that has to spill part of the model into system RAM will feel much slower than a smaller model that fits entirely in memory. When in doubt, choose the smaller quant that fits.
On Windows, check your graphics card with Task Manager or by running nvidia-smi in a terminal if you have an NVIDIA card. On a Mac, check Apple Menu, About This Mac for unified memory. Write the number down before you pick a download, because it decides everything that follows.
Step 2: Run it with Ollama, the fastest free route
Ollama is the simplest free path, especially if you are comfortable with a terminal. Download Ollama from ollama.com for Windows, macOS, or Linux, install it, then open a terminal and run:
ollama run qwen3.8:27b
That single command downloads the model on first use, loads it, and opens a chat prompt. The download is about 17 to 18 GB for the default build, so give it time and do not interrupt it. Once it finishes, you can type questions directly in the terminal.
Two options are worth knowing. Thinking mode is on by default, which makes the model reason before answering. For short, direct answers, turn it off with:
ollama run qwen3.8:27b --think=false
You can also switch thinking off inside a chat with the nothink setting. If you want longer documents or code files in context, raise the context length, for example by setting the OLLAMA_CONTEXT_LENGTH environment variable to 32768, but only if you have memory headroom. Longer context uses more memory and can slow generation.
On Apple Silicon, Ollama also offers MLX builds, which are often faster on Macs because they use unified memory directly. If you see an MLX variant for your Mac, try it when the standard build feels slow.
Step 3: Run it with LM Studio, the free graphical route
If you prefer buttons over a terminal, LM Studio is a free desktop app for Windows, macOS, and Linux. Install it, open the model browser, and search for Qwen3.8 27B. LM Studio shows which builds fit your hardware, which removes much of the guesswork.
Pick a 4-bit GGUF build if you have a 24 GB card or a well-equipped Mac, or a smaller quant if you have 16 GB. Download it inside LM Studio, wait for the progress bar to finish, then load the model and start chatting in the built-in window. You can adjust context length and temperature from the side panel without editing config files.
LM Studio is also a good way to test whether the model loads at all before you wire it into another app. If a build fails to load, try a smaller quant first, then check that LM Studio itself is up to date. Qwen3.8 uses a newer model architecture, and older runtimes can refuse to load it with an unknown architecture error. Updating the app usually fixes that.
Step 4: Give it a proper interface and a local API
A terminal chat is fine for testing, but most people want a nicer window or a way to call the model from their own code. Both are free.
For a chat interface, Open WebUI is a popular free, self-hosted option that talks to Ollama over its local server at localhost port 11434. Point Open WebUI at that address, select your downloaded model, and you get a clean web chat that still runs entirely on your machine. LM Studio also exposes its own local server, so check its server tab if you go the GUI route.
For code, Ollama exposes an OpenAI-compatible local API. Keep API keys out of this setup entirely: a local endpoint needs no key, and any tutorial that asks you to paste a cloud key into a local config is mixing two different setups. If your app supports a custom base URL, point it at localhost and select the Qwen3.8 model name exactly as Ollama lists it.
Step 6: Use thinking mode, vision, and context wisely
Qwen3.8-27B is more than a text chatbot. It supports image and video understanding, a native context of about 262K tokens, and adjustable thinking effort. Those features are powerful, but they all cost memory and speed, so use them deliberately.
For everyday chat, drafting, and coding help, a 32K context is plenty. Reserve very long context for genuinely long jobs, such as reviewing a large codebase, and expect slower output. For quick questions, turn thinking off; for maths, debugging, or multi-step planning, leave it on.
Vision surprises most local users: you can feed the model screenshots, diagrams, and documents without uploading them to a cloud service, which is ideal for private material. Keep image sizes reasonable, and test with one image before building a workflow around batches of files.
Troubleshooting the free setup
If the download stalls, check disk space first, free up room, then restart the download rather than forcing a partial file. If generation is extremely slow, the model is probably spilling into system RAM: drop to a smaller quant, shorten the context, or close memory-hungry apps. If the model refuses to load with an unknown architecture message, update your runtime. If Ollama reports that the model name does not exist, check the exact tag, because tags must match exactly.