13. The Qwen3.8 machine

Now we can put the pieces in one place.

Qwen3.8 is a decoder-only language model. Text in, next-token distribution out. The open checkpoint is a text model. The hosted Max product wraps that idea in vision, tools, and a longer default window.

It sits on the Qwen3.5 backbone, which itself sits on ideas Qwen shipped earlier as Qwen3-Next: hybrid attention, a very sparse MoE, multi-token prediction. It is not a bigger Qwen3-235B. Different skeleton.

The shape

The 92 layers are 23 repeats of this block:

DeltaNet → MoE
DeltaNet → MoE
DeltaNet → MoE
full gated attention → MoE

Inside each MoE: 512 experts, 10 routed plus 1 shared, expert width 2,048.

Full attention: 64 query heads, 4 key/value heads, head size 256, RoPE on 25% of the dims.

DeltaNet: 128 value heads, 16 query/key heads, head size 128.

There is also MTP — multi-token prediction. The model was trained not only to guess the next token but to guess a few ahead. At serving time this can feed speculative decoding: draft extra tokens cheaply, then check them. You can ignore this and the rest of the book still holds. It is a speed trick sitting on top of the same machine.

How it compares to what came before

Qwen3 is the last generation with a full public technical report. If you want the training recipe in lab-level detail, that paper is still the document. Qwen3.8's public materials are a model card, a config.json, a license, and a product blog.

Two products, one name in the press

People say "Qwen3.8 Max" for both of these. They should not.

The weights are Qwen3.8-2.4T-A95B (and an official FP8 quant). Thinking is mandatory. No images.

The API is qwen3.8-max. Images and video in, text out. Thinking can be dialed. Context is a million tokens. Function calling and structured output are on. Fine-tuning, on Alibaba's own table, is not.

There is supposed to be a dense Qwen3.8-27B for people with a single box. On the day the 2.4T files appeared, the official 27B repo was still a promise. Community repos with that name are not the official model.

What they say it is good at

The launch blog is about long jobs, not trivia.

They ran it for about 16 days on an empty repo and got a self-evolving coding harness, oh-my-cli, with hundreds of commits. They gave it a paper and GPUs and say it reproduced the experiments, then beat the paper's own method. They entered it in a human contest and say it beat most of the field in 24 hours. They describe a chip-design loop that cut a circuit from thousands of gates to 678, and a year-long shop simulation where it made money.

These are Alibaba's stories, on Alibaba's harnesses, graded in ways Alibaba chose. The published table is more mixed, and it is still a vendor table. Strong on PaperBench and computer-use style tests. Behind the current Anthropic flagship on some software-engineering benches. Treat "second only to Fable 5" as marketing until independent evals exist.

What you can take from the blog without buying the leaderboard: they trained and evaluated this thing as an agent in a loop. That matches the architecture. A 1M window, cheap long attention, mandatory scratchpads, and RL in real harnesses are all aimed at the same job — stay in a repo for a long time without losing the plot.

The license, briefly

Not Apache. Custom "Qwen3.8-Max License."

You can use, copy, modify, fine-tune, even sell, with two tripwires:

• If your product has more than 100 million monthly users, or $20 million monthly revenue, you have to show the model name in the UI.
• If you are in the business of selling model-as-a-service or a coding/office AI assistant, and your company does more than $50 million a year, you need a separate deal. Internal use that does not expose the model to third parties is carved out.

Read the file if you are going to ship something. The Hugging Face card is not a substitute for the LICENSE.