What 3B Active Parameters Actually Buys You: Qwen 3.6 35B A3B Explained

Model names have become a compressed spec sheet, and Qwen 3.6 35B A3B compresses harder than most. The 35B is a parameter count. The A3B is a second parameter count. People reasonably assume the second one replaces the first, and it does not.

Here is what the two numbers each control, and what changes for you depending on whether you are running the model on your own hardware or reaching it over an API.

Two numbers, two different costs

Every model has a memory cost and a compute cost, and for a long time they moved together. A dense model loads all of its weights and then uses all of them for every token it produces. One number described both.

A mixture of experts model splits that. The layers are divided into a set of specialised sub-networks, and a small router decides which few to send each token through. Qwen 3.6 35B A3B has 35 billion parameters in total, and roughly 3B of them fire for any given token.

So:

• 35B is the memory number. All the weights sit in memory the whole time, because the router can pick any of them for the next token and there is no time to go fetch one from disk mid-sentence.
• 3B is the compute number. That is how much arithmetic happens per token, which is what actually determines how fast text comes out.

The download does not shrink. The wait between words does.

Why this trips people up locally

At the 4-bit quantization almost everyone uses, this model is roughly 18 to 20 GB of weights, before the context window takes its own share. That puts the practical floor at a 24 GB graphics card, or a Mac with 32 GB of unified memory running at a slower token rate. On 8 or 16 GB it either refuses to load or crawls.

None of that follows from the "3B active" reading. Somebody who budgets memory for a 3B model is off by an order of magnitude, and the failure arrives as a load error or a swap-thrashing crawl rather than as a helpful message. Our guide on how much RAM you need for local AI works through where the memory actually goes.

The right way to hold it: this is a big model that behaves like a small one while it is talking. You pay for the big model in RAM and you get the small model's speed.

What you get for the trade

The reason to build a model this way is that quality tracks total parameters more than it tracks active ones. A dense 3B model is a small model in every sense, and it feels like one. A 35B model with 3B active has the breadth of the larger weight set to draw on and the response speed of something much lighter.

That is a good trade for two kinds of work. Long conversations, where a dense model of similar quality would leave you waiting on every turn. And agent loops, where the model produces a lot of intermediate output that nobody reads and the total wall-clock time is dominated by token throughput.

It is a worse trade if memory is your binding constraint. If you have 16 GB and you need something to fit, a genuinely small dense model fits and this does not, however few parameters are active.

The other two switches on this model

Two features matter as much as the architecture in day to day use.

Image input. The model reads screenshots, photos and diagrams alongside text. In practice this is the feature that turns it from a chat model into a debugging tool, because pasting the error dialog is faster than transcribing it.

A thinking mode you toggle. Reasoning is a switch here, not a fixed trait. That matters more than it sounds, because a reasoning pass is output tokens, and output tokens are what you pay for. Leaving Think on permanently means paying for working that nobody reads on questions that never needed it. Turning it on for a multi-step problem is money well spent. Hermes 3 405B, sitting in the same picker, has no thinking mode at all, which is a fair reminder that this is a per-model property rather than a platform setting.

Where the model sits in the catalog

Qwen 3.6 35B A3B is in the Hosted catalog, which is the €19 a month plan with 900,000 credits and 15 chat models. Its larger sibling, Qwen 3.6 27B, is Pro and Max only. That naming is genuinely confusing: the smaller-sounding number sits on the more expensive tier, which is backwards from what most people expect, and it is worth checking your tier before you go looking for it in the picker.

If you want to try the A3B without a subscription, a credit pack does that: €5 for 165,000 credits, €10 for 350,000, €25 for 900,000, any of them from your first purchase, and the credits never expire. As a rough chat reference, Hermes 3 70B in the same catalog costs 0.07 credits per output token, which makes the €5 pack around 2.3 million output tokens on that model. The plans and packs are here.

The reason a hosted route exists for a model this size is not that local is bad. It is that a 24 GB card is a strange thing to buy in order to evaluate one model. Running it on managed NVIDIA H100, A100 and B200 class GPUs and looking at it in a browser tab costs less than one evening of renting hardware, and it takes a minute instead of an afternoon. The details are on the run Qwen 3.6 without a GPU page, and the step by step version is in how to run Qwen 3.6 without a GPU.

The honest limits, stated once: shared hardware means a request can queue for a few seconds at peak, credits are one pool for text, images and video together, and the catalog is curated so you cannot bring your own base checkpoint. Account data is hosted in the EU and you can delete it yourself. We never train on your data and we never sell it. Content involving minors is refused everywhere, always.

The one-line version

A3B tells you how fast it answers. The 35B tells you whether it fits. Budget for the 35B, enjoy the 3B, and if the 35B does not fit, start on Hosted and let someone else own the card.