Qwen4-Max Local Download: Why It Doesn’t Exist and How to Use It

Qwen4-Max Local Download: Why It Doesn’t Exist and How to Use It

So I went looking for the download link. Hugging Face, Alibaba’s GitHub, the Ollama library, nothing. Qwen4-Max isn’t on any of them, and it took me a while to accept that it was never going to be.

A lot of posts right now call it the best local LLM, and I think that’s mostly keyword chasing. MarkTechPost describes it as a proprietary sparse mixture-of-experts model with a 2.4 trillion parameter backbone, where only 71 billion parameters are active for each token. It runs on Alibaba’s servers and nowhere else. That also means it’s not on Ollama, LM Studio or Hugging Face right now, and there’s no ollama run qwen4-max command. Any tutorial that shows one is guessing. Even if the weights dropped tomorrow, you wouldn't run 2.4 trillion parameters on a laptop, and a quantized build of something that size is still a server-room problem.

The local Qwen 4 model exists, but only on a slide. Alibaba announced a 27B open-weights tier at Apsara in September, and as of today it has no specs, no price and no date.

What Qwen4-Max actually is

It’s the flagship of Alibaba’s Qwen 4 generation. It opened as a two-week preview on November 11 under the name qwen4-max-preview, only through the Token Plan subscription, and now it’s listed as a normal API model called qwen4-max on QwenCloud. The context window is 1 million tokens. It takes images and video as well as text, but it only answers in text. Thinking is on by default, and it supports tool calling, web search, JSON schema output and prefix completion, where you give it the start of an answer and it continues strictly from there. Alibaba pitches it for coding, office work, frontend design and long agent tasks. One thing it can’t do, at least on Alibaba’s Model Studio page: no fine-tuning, no batch inference, no deploying your own copy.

And then the max output number. QwenCloud says 393K tokens. Alibaba’s Model Studio page says 262,144 with thinking on and 131,072 with thinking off. Two official pages, two different numbers, and I couldn’t find anything that explains it. My guess is one page counts the reasoning budget and the other doesn’t, but that’s only a guess, so test it yourself before you design anything around the 393K figure.

Pricing is simple enough. Input is $1.5 per million tokens, output is $4.5, and cached input reads cost $0.075. Say you send a 50,000-token codebase and get 5,000 tokens back. That’s about ten cents. Send the same codebase again within the cache window and the input part drops to less than half a cent. Worst case, one call that uses the whole 393K output budget would cost around $1.77, and it’s worth knowing that thinking tokens share that budget with the answer.

Alibaba also says it cut API prices and that average cost per task fell 62%. That’s their number, not mine, and I haven’t seen anyone reproduce it.

On quality, I haven’t run it on my own work yet, so everything here is other people’s testing. MindStudio did a hands-on and found it overthinks less than Qwen 3. It also caught its own labeling mistake on a map, and it spotted an improperly worn seatbelt strap in a photo that Qwen 3 reportedly missed. In another test it turned a long written brief into working front-end code for an animated scene. These are single tests, so treat them as hints.

The benchmark problem

Nobody has run it independently yet.

What exists is vendor-reported. Gigazine, as relayed by the Italian site ilsoftware, says Qwen4-Max beats Claude Opus 5.5 and GPT-6 Astra on some tests, including DeepSWE v1.1 for coding agents. “Some tests” is doing a lot of work in that sentence. Meanwhile BenchLM, a tracker that only compares models when public results exist on both sides, lists Qwen4 Max as not ranked, with zero public rows against 17 for Opus 5.5 on one lane. Orcarouter noticed the same gap from the other direction: for SWE-bench Verified, SWE-bench Pro and Terminal-Bench, the competitor has numbers and Qwen4 Max has none reported.

There’s a pattern from earlier Qwen releases that I’d keep in mind. A comparison on danilchenko.dev put Qwen 3.7 Max at 60.6% on SWE-bench Pro and 69.7% on Terminal-Bench 2.0. Claude Opus 4.8 scored 69.2% and 74.6% on the same two. Qwen did win on LiveCodeBench in that table, with 91.6%, but there was nothing to compare it against. So the Max models tend to look great on the launch slides and a bit less great once somebody else runs the harness.

I’m putting my money on 70/30 here: it’ll be good, probably very good at frontend work, and it won’t beat everything. Wait two weeks for independent numbers before you move a production workload.

Using it through the API

You don’t need to learn a new SDK. QwenCloud is compatible with the OpenAI and Anthropic API formats, so the normal OpenAI Python package works. The setup will only take a few minutes. Make an account, create an API key in the console, and copy the base URL for your region from the same page (the Model Studio docs I read were for the Singapore international region). I’m not pasting a URL here because I’d rather you copy it from the console than trust mine.

import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["QWEN_API_KEY"],
base_url="PASTE_BASE_URL_FROM_YOUR_CONSOLE",
)
resp = client.chat.completions.create(
model="qwen4-max",
messages=[
{"role": "user", "content": "Review this function and list bugs."}
],
)
print(resp.choices[0].message.content)

Use qwen4-max, not qwen4-max-preview. The preview ID was only supposed to live until November 25, and bills reportedly show an alias, qwen4.0-max-preview, which is confusing if you're matching invoices.

If you don’t want to write code, the same model is in Qwen Studio and the Qwen app, and early access was also opened in QwenWork and Qoder. For a first test I’d skip all the fancy stuff. Paste in the longest, ugliest document you have and see if it keeps track of details from page 1 when it reaches page 300. That’s what a 1M window is supposed to be for, and it’s the one claim you can check in ten minutes.

What you can run at home today

If what you wanted was a model on your own machine, the one I’d look at is Qwen3.8–27B, from the previous generation. It’s open weights, takes images and video, has a 256K context window, and switches between thinking and non-thinking modes. One quantizer’s model card on Hugging Face says the license is Apache 2.0, inherited from the original. Check the original card yourself before you ship anything.

First update Ollama. This model needs version 0.32.12 or newer, and older installs will throw an error on the pull. Open terminal and run these two commands:

ollama pull qwen3.8:27b
ollama run qwen3.8:27b

The download itself is around 18GB for the default Q4_K_M build, so start it before you go make tea. A tester at Computing for Geeks put it on an RTX 4090, where the model loaded in 7.6 seconds and sat at 19.7 GiB of VRAM with Ollama’s default 32K context, giving roughly 100 tokens per second. That’s quick enough that chat feels instant. Apple Silicon works too, if you have about 24GB or more of unified memory. A 16GB card can technically load it, but it’ll crawl.

Context is where it bites. On that same 4090, pushing the context to 131K with the default cache made 30% of the model spill onto the CPU, and speed fell to 16.4 tokens per second. Quantizing the KV cache to q4_0 got 131K to fit on the card. So a 24GB GPU is fine for chat and normal coding help. It isn’t fine for dumping a whole repo into one prompt without tweaking settings.

Ollama also serves an OpenAI-compatible endpoint on localhost:11434, which means the Python code above works with a one-line change. Swap the base URL, set the model name to qwen3.8:27b, and you can compare the local model and Qwen4-Max on the same prompts. That's the most useful test I can think of.

The giant sibling, Qwen3.8–2.4T-A95B, did get released as weights too, under a custom license. At 2.4 trillion parameters it’s too big for any home machine, so skip it.

One more name you’ll see in the Ollama library is qwen3.8-flash-next. Its page calls it an experimental preview of the architecture Qwen4 will use, with 125B total parameters, 6B active per token, and a smallest download of 105GB. It's not Qwen4, and it's too big for most home machines anyway, so don't pull it expecting a small Qwen 4 model.

Where I land

Qwen4-Max is a hosted model, and the honest answer to “is it the best local LLM” is that it isn’t local at all. Whether it’s the best model overall is open until someone besides Alibaba runs the tests.

For most people I’d say start with Qwen3.8–27B on a 24GB card and use the Qwen4-Max API only for the jobs that need a million tokens of context or the best frontend output. That’s too broad, actually. If your work is long agent runs over a large codebase, no 27B is going to keep up, and you should just pay the few dollars.

One more loose end. In the 3.8 generation Max came first, the big weights followed about ten days later, and the 27B and Flash tiers landed within a month. Alibaba has reportedly said it plans to publish an open-weight version of the Qwen4 family, but I only found that in one secondary source, and there’s no date. Until the weights show up on Hugging Face and then in the Ollama library, Qwen4-Max stays API-only. If it does, the 27B download link I couldn’t find on day one might show up before the end of the year. I’d bookmark the Qwen page on Hugging Face and check it after November 25, when the preview window closes.

Post a Comment

Previous Post Next Post