If you have been eyeing up the Qwen 3.8 family and wondering which size to actually download, here is my straight answer: Qwen 3.8 27B is the build most people should run. It is big enough to power a proper AI agent, small enough that a well-specced single machine can run a quantised copy, and once the download finishes it is yours for good. In this guide I will cover what it is, how to run it locally, and how to turn it into a working agent brain.
I have been testing the Qwen 3.8 generation hands-on this week, so none of this is secondhand.
📺 Watch: I Tested Qwen 3.8 Max So You Don't Have to...
What Qwen 3.8 Is, and Where the 27B Sits
Qwen is Alibaba's family of open-weight models, and it has become one of the most respected open lines going — known for strong coding, reasoning and multilingual performance. Qwen 3.8 is the current generation, and like its predecessors it ships in a spread of sizes: small distilled builds at the bottom (community 1.2B and 4B distills are circulating), the hosted flagship Qwen 3.8 Max at the top, and the 27B sitting in the useful middle. My how to use Qwen 3.8 guide has the full family tour.
Open-weight is the part that matters, because it changes the ownership model entirely:
- The model files are published for anyone to download.
- Once downloaded, the model runs offline on your machine.
- Nobody can rate-limit you, change the terms, or take the model away.
- You can swap it in and out of your tools as freely as you like.
Why Qwen 3.8 27B Is the Sweet Spot
The tiny builds are quick and run on almost anything, but they are limited — fine for small jobs, frustrating as the main brain behind a serious agent. Qwen 3.8 Max is the strongest expression of the family, but it is hosted: brilliant, and never actually yours.
The 27B class splits the difference. It is big enough for real agent work — multi-step reasoning, tool use, longer writing tasks — yet small enough that one well-specced machine can run a quantised build. That balance is exactly why, when people ask me about the best free AI model for a Hermes agent, the 27B class is where I point them first.
📺 Watch: China's NEW Qwen 3.6 27B is INSANE (FREE + OpenSource!)
Quantisation in Plain English
Full-precision weights for a model this size are heavy, and that is where quantisation earns its keep. Quantisation stores the model's numbers in a more compact format, trading a little quality for a lot less memory — and that saving is what makes Qwen 3.8 27B realistic on ordinary hardware in the first place.
You do not need to do any of it yourself. The community has already published quantised versions of Qwen3.8-27B on Hugging Face — FP8 and NVFP4 formats among them — built specifically for running the model efficiently on your own machine. Pick the version that fits your memory; the hard part is already done.
How to Run Qwen 3.8 27B Locally
Here is the route I use, the same one from my free local install guide:
- Install Ollama. Free, runs on Mac, Windows and Linux, and handles the awkward parts of running models locally.
- Pull a Qwen 3.8 build that fits your machine. A quantised 27B if your memory allows, a smaller build if not. The first download is several gigabytes, so let it finish before judging anything.
- Run the model. Once it loads, you are talking to Qwen entirely on your own hardware — no account, no meter, no internet required.
- Prefer something visual? LM Studio is the alternative — a friendly interface for downloading, testing and chatting with local models if the terminal is not your thing.
- Use the local endpoint. Either way, you end up with a local endpoint on your machine that any agent tool can plug into — and that endpoint is what turns a chat model into an agent brain.
If you want a local AI agent running without the trial and error, the AI Profit Boardroom has the Agent OS with local-model profiles set up for you. → Get the done-for-you local setup
📺 Watch: NEW Qwen3.6-27B Open Source Update is INSANE (FREE!)
Wiring It Into the Hermes Agent
A model on its own just chats. The interesting part starts when you give it a body, and for me that body is Hermes. Hermes is provider-agnostic — it does not care whose model it runs on. Cloud, local, big, small: they all slot into the same place.
With your model running locally, point Hermes at it. Run the hermes model command and select your Qwen build from the list, or set it from the models page in the Hermes web UI. That is genuinely the whole job. I walk through the pairing in detail in my Hermes agent with Qwen 3.8 guide, and the general process in the Hermes local model setup walkthrough.
The framing I always come back to: Hermes is the body — the tools, the file access, the ability to actually do things — and Qwen 3.8 27B is the brain. Same body, swappable brains, and this particular brain lives on your own machine. It is exactly how I run my own Agent OS: one agent framework on top, with models slotted in and out underneath depending on the job.
The Big-Plus-Small Pattern
Here is the pattern that has proven itself across my setups: pair a larger Qwen for the heavy reasoning with a smaller one as a fast sub-agent. The 27B does the thinking — planning, writing, working through multi-step problems — whilst a small distill handles the quick, repetitive jobs like summarising files and triaging inputs. In my own setup that means the 27B carries the agent while a 4B distill runs point on the small stuff.
Why bother? Resources. Every task you hand the small model is memory and time the big one does not burn, and agents fire off far more small tasks than big ones, so the saving compounds all day. Big brain for the hard calls, small brain for the busywork: that is how one machine starts to feel like a team.
Qwen 3.8 27B Local vs Qwen 3.8 Max Hosted
| Qwen 3.8 27B (local) | Qwen 3.8 Max (hosted) | |
|---|---|---|
| Where it runs | Your own machine, fully offline once downloaded | Alibaba's servers, accessed over the internet |
| Cost | Free to run — hardware and electricity are the only costs | Paid usage — the meter runs whilst you work |
| Control | Total — no rate limits, no terms changes, nobody can switch it off | You are on their infrastructure, on their terms |
| When to use | Daily agent work, private data, always-on automations | One-off heavy tasks where you want the family's strongest output |
What My Qwen 3.8 Testing Says
I have spent this week inside the Qwen 3.8 generation. I published "I Tested Qwen 3.8 Max So You Don't Have to..." on YouTube in early August, plus a comparison running Qwen 3.8 Max against Fable 5 and GPT-5.6 Sol. The way I judge these models never changes: Goldie Bench — my own side-by-side runs on real tasks, actual agent jobs and actual content work rather than leaderboard screenshots. That is how the Qwen family gets measured against the frontier models in my videos, and if you have seen them you will know I do not hand out praise easily.
The honest takeaway from that testing: this family is genuinely strong. Max is the strongest expression of it, and it held its own in serious company. But Max is hosted — you rent it. The 27B is the same family in a version you can own, and for daily agent work that difference is the whole story.
Qwen 3.8 27B FAQ
Can my machine actually run it?
It mostly comes down to memory. The quantised community builds exist precisely to pull the requirements down, so a well-specced modern machine — plenty of RAM or a decent GPU — has a genuine shot at the 27B. If it will not fit, step down to a smaller Qwen build, or use the free API route for Hermes whilst you plan a hardware upgrade.
Is it really free?
Yes. The weights are open and the model runs on your hardware, so electricity is the only running cost. No subscription, no tokens, and no anxiety about a long agent run chewing through a usage cap at two in the morning.
Should I run the 27B or use Qwen 3.8 Max?
If you can only pick one, own rather than rent. Run the 27B for everyday agent work, and treat Max as the occasional big gun for tasks where you want the family's absolute best and do not mind it living in the cloud.
Does it work with Hermes?
Yes — Hermes is provider-agnostic, so a locally running Qwen build shows up as just another model to select. Run it through Ollama, pick it in Hermes, and you are away.
What about the small distills?
The community 1.2B and 4B distills are brilliant for what they are: sub-agents, quick tasks and older hardware. I just would not make one the main brain of a serious agent. Start small if you have to — but the 27B is where the family gets properly capable.
My Verdict on the 27B
The 27B is my pick for most people for one simple reason: it is the point where the Qwen family stops being a toy and starts being infrastructure. Download it once, run it through Ollama, point Hermes at it, and you have an agent brain that costs nothing to run and answers to nobody. Rent Max for the peaks if you like — but own your baseline.
If you want an AI agent running on a model you actually own, check out the AI Profit Boardroom — inside you get the Agent OS, model setups like this done for you, daily tutorials, weekly live coaching calls, and me answering your questions personally. → Set up your own Qwen-powered agent
