How to Run Qwen 3.7 Locally with Ollama (2026)

You searched for how to run Qwen 3.7 locally, so here is the honest answer up front: you cannot, not yet. Qwen 3.7 Max and Plus are closed-weight, API-only models, and Alibaba has not released open weights for the 3.7 line. The good news is twofold. The newest open Qwen, version 3.6, is excellent, runs on your own machine today, and installs with one command through Ollama. And Alibaba's pattern is to follow each API launch with open weights about three to four weeks later, so Qwen 3.7 weights are expected around mid-2026, and when they land this exact guide works by changing one word. So let us get a real, private, free Qwen running now, and set you up to swap to 3.7 the day it drops.

The cloud version of Qwen 3.7 is genuinely strong, and if you only need an API, our Qwen 3.7 Max vs GPT-5.5 comparison covers it. But running a model on your own machine is a different goal: free generation with no meter, complete privacy, and it works on a plane. The catch, as above, is that the 3.7 line is closed for now. So this guide does the useful thing, it gets the newest open Qwen running locally today, in a way that becomes a 3.7 guide the instant Alibaba ships those weights. No theory, just the commands.

Can you run Qwen 3.7 locally right now?

Short answer, no, and it is worth being clear why. Alibaba ships its flagship tiers (Max and Plus) as closed, API-only models first, then releases open weights for the line a few weeks later. Qwen 3.6 followed exactly that path. So Qwen 3.7 open weights are not a question of if but when, and the pattern points at the middle of 2026.

In the meantime, Qwen 3.6 is the newest open-weight Qwen, and it is not a consolation prize. It ships with vision (text and image input), a 256K-token context window, and strong agentic coding, all under the permissive Apache 2.0 license. It is on Ollama right now. Everything below uses 3.6, and the only thing that changes for 3.7 is one word in one command.

The fastest path: Ollama

Ollama is the quickest way from nothing to a running model. It bundles the inference engine, a model library and a local API behind a single command. Install it, then pull and chat with a model in one line.

# macOS and Linux (Windows has a normal installer at ollama.com)
curl -fsSL https://ollama.com/install.sh | sh

# pull and run the newest open Qwen, then just type at the prompt
ollama run qwen3.6:27b

That second command downloads the model the first time (about 17 GB for the 27B), then drops you into a chat prompt. Type a message, or /bye to leave. The model stays on disk, so the next launch is instant and fully offline.

Pick the right size for your hardware

The single most common mistake is pulling a model too big for your machine, which either fails or crawls. Match the model to your VRAM (or to unified memory on Apple Silicon), at the default Q4_K_M quantization. Qwen 3.6 is a large model, so for lighter machines the earlier but still excellent Qwen 3 dense models are the better fit.

A good rule: the 8B is the sweet spot that punches above its weight for everyday coding and chat, the 14B is a clear step up in reasoning, and the Qwen 3.6 27B is where you feel near-frontier quality if your hardware can hold it. When in doubt, start one tier below what you think you can run and move up.

Quantization, in one minute

Quantization shrinks a model's weights to fewer bits so it fits in less memory. You will see tags like Q4, Q5, Q6 and Q8. The one to use is Q4_K_M: it cuts memory by roughly half versus the full model while losing under one percent on benchmarks, which is why Ollama defaults to it. Go higher (Q6, Q8) only if you have memory to spare and want the last sliver of quality. If you run long contexts and run short on memory, quantizing the KV cache as well buys back a surprising amount of room.

Other ways to run it

Ollama is the easy default, but it is not the only door.

• LM Studio is a polished desktop app with a visual model browser and a chat window. If you would rather click than type, start here; it runs the same GGUF models.
• llama.cpp is the engine underneath most of these tools. Use it directly when you want maximum control, custom flags, or to run a specific GGUF from Hugging Face with huggingface-cli.
• vLLM and SGLang are for production serving, many users, high throughput. They need a CUDA GPU, so on a Mac you stay with Ollama or LM Studio.

For most people reading this, Ollama for daily use plus LM Studio when you want a GUI covers everything.

Use it from your code

This is the part that makes a local model genuinely useful: Ollama serves an OpenAI-compatible API at http://localhost:11434/v1. So your existing OpenAI client works by changing the base URL and the model name.

from openai import OpenAI

client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")
resp = client.chat.completions.create(
    model="qwen3.6:27b",
    messages=[{"role": "user", "content": "Refactor this function"}],
)
print(resp.choices[0].message.content)

That is the whole integration. Because the cloud Qwen and OpenAI also speak this format, you can point the same code at a local model, a Qwen API, or another provider by changing two lines, which makes testing local against cloud almost free.

When Qwen 3.7 open weights arrive

Here is the payoff for doing this now. When Alibaba publishes the Qwen 3.7 open weights, Ollama will add a qwen3.7 tag within days. At that point you change exactly one thing:

ollama run qwen3.7

Your install, your size logic, your quantization choice and your code all stay the same. So set up 3.6 today, get comfortable, and the upgrade is a single pull. If you want, keep an eye on the Ollama Qwen library for the tag to appear.

Sources and further reading

• Ollama, Qwen 3.6 model library
• Ollama, Qwen 3 model library
• How to run Qwen 3.7 locally, the honest 2026 answer
• Qwen docs, running locally with llama.cpp

Frequently asked questions