Qwen 3.8 Max vs A95B: which one, and what for

Qwen 3.8 arrived as a family, and two of its flagships are in the LU Labs Cloud chat catalog under names that look like a ladder: Qwen 3.8 Max and Qwen 3.8 A95B. People reasonably read that as a smaller one and a larger one, or as an older one and its successor.

Neither reading is right. They sit side by side because they are different handling of the same class of model, and the thing that separates them in daily use is the reasoning switch.

What A95B's name is telling you

A95B is the open weight checkpoint: 2.4 trillion parameters in total, roughly 95 billion of them active for any single token, with a context window of 262,144 tokens.

Those two parameter counts do different jobs, and conflating them is the most common mistake with any mixture of experts model. The total is a memory number: all of it sits loaded, because the router can send the next token through any part of it and there is no time to fetch a piece from disk mid-sentence. The active count is a compute number: that is how much arithmetic actually happens per token, which is what determines how fast the text comes out.

So A95B is a very large model that behaves like a much smaller one while it is talking. You pay for the large model in hardware and you get the small model's throughput. On our side that hardware is already paid for, which is the entire reason this page exists rather than a download link.

On the licence: the weights are public, but they are not Apache 2.0. The repository carries Alibaba's own terms. The Apache 2.0 model in that same release wave is the small dense 27B, which is a different model, and worth keeping straight if licensing matters to your project.

What Max is

Max is the other flagship, and the useful facts about it are behavioural rather than architectural. It is the cheaper of the two per token in our catalog, it reasons by default, and its reasoning can be switched off.

We are deliberately not going to tell you its parameter count or its internal layout. Those are not published in a form we can stand behind, and a spec sheet assembled from guesses is worse than no spec sheet.

Neither model is the other's successor. Max did not replace A95B and A95B did not replace Max. They are both in the picker because they are both worth having, and there is no upgrade path implied between them.

The difference that actually decides it

Both models reason before they answer. On Max you can stop that. On A95B you cannot.

This is not a preference setting that we chose to expose on one and hide on the other. It is what the two upstreams accept. Max answers to the chat template's own thinking flag, and turning it off takes a short question from around a hundred output tokens down to a handful, with the same answer at the end. A95B refuses the equivalent request outright: asking it not to think comes back as an error, not as a quieter answer.

There is a second layer worth knowing, because it catches people who assume every reasoner works the same way. The reasoning effort control that moves the GLM 5.3 pair does nothing on Max. Sending a lower effort rung does not shorten its reasoning, and on some runs it produced slightly more. The lever on Max is the Think button and only the Think button.

What that costs

Credits translate directly from the catalog price, at one credit for $0.00001.

	Input, per 1M tokens	Output, per 1M tokens
Qwen 3.8 Max	165,000 credits	495,100 credits
Qwen 3.8 A95B	200,000 credits	600,000 credits

Max is about 17 percent cheaper per output token before you touch anything. Then the switch compounds it: on work that does not need a reasoning pass, turning thinking off on Max removes most of the output tokens from the turn, and A95B has no equivalent saving available at any price.

Both are in the catalog every plan and every credit pack carries, so the question is only how many credits you hold. A €5 pack of 165,000 credits spent entirely on Max is roughly 333,000 output tokens, and roughly 275,000 on A95B. A Pro month of 2,350,000 credits is roughly 4.7 million and 3.9 million respectively. No figure here survives contact with a real week, because input tokens draw from the same wallet and a long conversation resends its history on every turn.

Where they are the same

Both call tools through the standard parameter, natively, so agent scaffolding needs no translation layer on either.

Neither accepts image input. The upstream refuses a picture on both, which surprises people who have read that the Qwen family is multimodal. Family names predict nothing here; each model id has to be checked on its own.

Both are reachable from the same personal API key at https://lu-labs.ai/api/inference/v1 in the OpenAI chat completions shape. The only thing that changes in your code is the model id.

How to pick

Default to Max. It is cheaper per token, and the off switch means you can stop paying for reasoning on the two thirds of your prompts that never needed it. For extraction, reformatting, translation and summarising, Max with thinking off is not a compromise, it is the correct tool.

Reach for A95B when you want a model that always works the problem before it answers and you do not want the option of accidentally skipping that. Long derivations, code that has to be reasoned about rather than pattern matched, analysis where a fast plausible answer is worse than a slow correct one. The mandatory reasoning is a feature in that context and a bill everywhere else.

And if what you actually wanted was a Qwen 3.8 on your own hardware, that is the 27B, and it exists: how to run Qwen 3.8 27B on your own computer. To get either flagship running in a browser instead, start at run Qwen 3.8 without a GPU.