Qwen 3.8 is free. The "small" file is 397GB

Yesterday I covered Meta's new model that fits on a single gaming GPU. Today the exact opposite dropped: Alibaba released the open weights for Qwen 3.8, a 2.4 trillion parameter frontier-class model, free to download right now. The catch is the size: even the smallest, most aggressively compressed version is a 397 gigabyte file that wants about 450 gigabytes of RAM. This one was new territory for me too, so this letter is written the way I learned it: in plain words.

What actually shipped

About a week ago Alibaba promised open weights for Qwen 3.8 "next week." The much-anticipated 27B version slipped (it's still missing, the links 404), but the big one landed: Qwen3.8-2.4T-A95B, with Unsloth announcing ready-to-run builds the same day.

Unsloth, if you haven't run into them, is a tiny and famous team in local AI. When a new model drops, they're the folks who shrink it and package it so normal computers can load it, usually within hours. Think of them as people who vacuum-pack shipping containers into suitcases. Sometimes several suitcases.

Decoding the name, in plain words

2.4T-A95B means 2.4 trillion parameters total, with only 95 billion active at once. That's a mixture of experts: picture a hospital with thousands of specialists on staff, where only a handful walk into the room for any one patient. It's why a model this size can answer quickly. The trap is that you still had to build the whole hospital: all 2.4 trillion parameters have to live in memory, even though only 95 billion work at a time.

GGUF is just a file format, the MP3 of local AI models. It's the packaging that apps like Ollama and llama.cpp know how to open. When someone posts "the GGUFs are up," they mean the run-it-at-home version exists.

The sizes, per Unsloth's guide: lossless is 4.9 terabytes. The 4-bit build is 1.3TB, 2-bit is 657GB, and the extreme 1-bit build is 397GB, a 91% reduction. Even that smallest one wants roughly 450GB of RAM, and the rule of thumb is RAM plus VRAM should equal the quant size, or it falls back to slow disk offloading. Context window: 256K native, extensible to about a million tokens.

Reality check: who can actually run this?

My machine is a beefy AI box with 128GB of unified memory, and the smallest version of this model needs roughly three and a half of my entire computer. Realistically this is for used server hardware with 512 to 768GB of RAM, enterprises with compliance reasons to keep AI in-house, researchers, and the extreme end of the homelab crowd. Even then, Unsloth's own numbers suggest around 20 tokens per second on serious datacenter GPUs, which is not fast.

This is Kimi K3 territory: free is free until you price what it takes to use the free thing. You're not being handed a free boat, you're being handed a free yacht. The yacht is real. So are the docking fees.

Two honest cautions. First, 1-bit compression is extreme, and nobody knows yet how much intelligence survives that level of vacuum-packing; the benchmark numbers on the card are vendor numbers, day one, with no independent testing. Second, the license is Qwen's own custom license, not Apache 2.0 like Meta's Glimmer yesterday, so this is open weights, not open source; if you build on it, find and read the terms.

Why a file almost nobody can run still matters

Four days on this channel: Kimi K3 escaped a test sandbox, Meta shipped a one-GPU agent model, and now Alibaba published a frontier-class model's actual weights. These aren't separate stories. It's one story.

Training these models costs billions. The moment the weights go public, that multi-billion-dollar artifact becomes a free file on Hugging Face. So what exactly are the closed labs selling? If a free file gets you 90% of what a $200-a-month subscription does for your specific use case, the money has to move somewhere: speed, hosting, tools, trust, enterprise support. At a certain point it stops being about the model. I think that's why Meta suddenly rediscovered open source this week, and why the pressure lands on everyone charging per token.

To be fair, the giants aren't sitting still. Frontier closed models are still clearly ahead, and serving AI at scale is its own fortress. It's too early to call winners. That's exactly why I'm documenting this as it happens.

What to watch

Two things. The Qwen 3.8-27B, the version that will actually fit on machines like mine, is still "coming soon." When it lands I'll cover it and run comparisons. And the independent benchmarks of this 2.4T release will tell us whether the 1-bit magic trick preserved the brain.

Meanwhile, as I recorded this, my computer was literally running a 10-test tournament between four local models, including Meta's Muse Glimmer from yesterday. That comparison is the next video.

Sources

I am documenting a life of interest and curiosity, tech, money, health, and family included. New video daily.