What's hidden in a model name?

A35B is the part that works

In Qwen3-Coder-480B-A35B-Instruct, A35B stands for 35 billion active parameters. The model holds 480 billion parameters in total, the learned numbers it is made of, and that is the 480B just before the A. It reads and writes text in tokens, small pieces of words, and only 35B of its parameters do any work on each token. Both numbers are on Qwen's model card.

The A shows up because the model is a mixture of experts (MoE), built from many small parts of which only a few work on each token. Each of the two numbers then costs you something different. The total sets how much memory the model needs, because every part has to be loaded and ready. The active number sets how much work each token takes, which drives speed and running cost. That split is how DeepSeek-V4.1-Flash, a 552B model, works with only 8B for each token it reads.

Five families come up again and again in vendor decks: Qwen, DeepSeek, Gemini, GPT and Claude. Qwen and DeepSeek publish their model files and sizes, Qwen in the name itself and DeepSeek on the model card. Claude, GPT and Gemini publish no sizes at all, so their names follow other rules.

How do you read a model name?

Split the Qwen name at the hyphens and you get five parts, each answering a different question.

Qwen3 is the family and its generation, and Coder the specialty, since this model was built for code. The two sizes follow, 480B in total and A35B active. Instruct at the end describes the tuning, here for following instructions. Qwen adds that this model runs only in what it calls non-thinking mode, so it answers without first working through the problem.

Sizes don't stop at billions. Qwen3.8-2.4T-A95B holds 2.4T parameters, T for trillion, which is 2,400B, and uses 95B of them per token.

A name without an A part needs a second look. Qwen3.6-27B has none because it is a dense model, which uses all 27B of its parameters on every token. OpenAI's open-weight gpt-oss-120b has no A part either, yet it is a mixture of experts with 117B in total and 5.1B active. Even the 120 in its name is a rounded number next to the 117B on its card. When the name leaves something out, read the model card.

A tag like FP8 or BF16 gives the precision, how many bits each parameter is stored in: 16 bits (2 bytes) for BF16 and 8 bits (1 byte) for FP8. Qwen3.6-27B-FP8 is Qwen3.6-27B stored at 8 bits, so it needs less memory and still has all 27B parameters.

A number at the end can also be a date. DeepSeek released V4.1-Flash on 10 September 2026 and an update with the same architecture on 17 October, named V4.1-Flash-1017 after the month and the day. Tier words such as Flash, Lite, Pro and Mini are a different matter, because each maker decides what they mean.

Why does 552B run like 8B?

Inside an MoE model, most of the parameters sit in many small networks, the experts. For every token, a router picks a few of them, and only those are active while the rest wait in memory. The output of the active experts carries on through the model toward the answer.

DeepSeek-V4.1-Flash has 384 routed experts. Its router picks 6 of them for each token, and one shared expert joins every time. DeepSeek puts the result at 8B active parameters while the model reads your input and 16B while it writes the answer, out of 552B in total. Qwen3.6-35B-A3B works the same way on a smaller scale, with 8 of its 256 experts plus one shared expert per token, which comes to 3B of 35B.

Put six models side by side and every MoE among them keeps its active part under 9% of its total. DeepSeek-V4.1-Flash reads with 8B of 552B, about 1.4%, the smallest share in the figure. Qwen3.8-2.4T-A95B and gpt-oss-120b land close together at 95B of 2.4T (4.0%) and 5.1B of 117B (4.4%). Qwen3-Coder-480B-A35B-Instruct runs on 35B of 480B (7.3%), and Qwen3.6-35B-A3B on 3B of 35B (8.6%). Only the dense Qwen3.6-27B uses everything, 27B of 27B.

Memory works differently. The router can pick any expert for the next token, so all of them have to stay loaded. For the weights alone, a quick estimate is total parameters times bytes per parameter, which comes to about 2 GB per billion parameters at BF16, 1 GB at FP8 and 0.5 GB at 4 bits. The context and the serving software need memory on top of that.

Take the two Qwen3.6 models. Qwen3.6-35B-A3B at BF16 needs about 70 GB for its weights (35 × 2), even though only 3B are active per token. Qwen3.6-27B-FP8 needs about 27 GB (27 × 1) and uses all 27B on every token. Part of that gap is precision, but even stored the same way the MoE is the larger one, 35B against 27B, and it still does less work on each token, which is how it can be quick to answer.

What do Flash, Sonnet and Luna mean?

Claude, GPT and Gemini are API-only, so you reach them through the maker's service and never get the model files. Their names carry a generation number and a tier word that marks a size or speed level inside one family, and only the maker's own description tells you what a tier is for.

The table gives each maker's own description, from Anthropic's Claude models overview, the OpenAI models page and Google's Gemini models page. Names and numbers in this post are as the makers published them on 21 October 2026, and these families update often. Prices change often too, and each maker publishes them on its own pricing page.

Anthropic also rates the four Claude models against each other for speed, with Fable slower, Opus moderate, Sonnet fast and Haiku fastest. All four take text and images. If you are weighing Claude Opus vs Sonnet, Anthropic suggests starting with Opus 5.5 for most workloads.

For the context window, how much text a model can take in at once, the makers list 1M tokens for all four Claude models, Gemini 4 Argon and Gemini 3.8 Flash. OpenAI lists 1.05M for its three flagship models, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. Google's Gemini 3.8 page adds that Gemini 3.8 Flash writes up to 64K tokens of output and has thinking levels you can set to low, medium or high.

OpenAI describes GPT-6.1 Sol as near-Astra even though GPT-6 Astra carries the lower number, so a higher generation number does not mean a higher tier.

Watch the word Flash, because two makers use it for different things. DeepSeek's Flash is a 552B open-weight model, which DeepSeek calls "the smallest model in our new architecture family, with native visual understanding". Google's Flash is a fast tier of a closed family, and its size is not published. So when a vendor says they run Flash, ask whose.

Which models can you run yourself?

The Qwen, DeepSeek and gpt-oss models in this post come as open weights, so you can download the model files and run them on your own hardware under the license that comes with them. Open weights don't always mean self-hosting, though, since DeepSeek also serves V4.1-Flash through its own API under the name deepseek-flash.

Start with the license. DeepSeek-V4.1-Flash comes under MIT, and gpt-oss-120b and three of the four Qwen models under Apache 2.0. The largest Qwen here, Qwen3.8-2.4T-A95B, has its own Qwen3.8-Max license, so even one family can carry different terms. If you are comparing Qwen vs DeepSeek for your own servers, read the license before anyone orders hardware.

Then size the hardware with the memory estimate. DeepSeek's published files include weights stored at 8 bits (FP8), and at 1 byte per parameter, 552B comes to about 552 GB for the weights alone. That means a server with several GPUs, for a model that works with 8B to 16B per token. OpenAI stores the expert weights of gpt-oss-120b in a 4-bit format called MXFP4 and says this lets it "run on a single 80GB GPU".

Inputs and answer style differ too. DeepSeek-V4.1-Flash takes text and images, while Qwen3.8-2.4T-A95B is text-only and always works through a problem before answering, in what Qwen calls thinking mode. Qwen3.6-35B-A3B does that by default, and you can switch it to answer directly.

Which model fits which job?

A customer service operation gives a model very different jobs. A live call needs a reply while the caller is still on the line, and nobody is waiting on the summary of yesterday's calls. The table follows the makers' own descriptions, so read each row as a place to start testing rather than a ranking.

The live call is the strictest row. A caller hears every pause, and the model is only one source of it; our post on where the wait on a voice line comes from splits that pause into parts you can measure.

Pick two or three candidates from the row that matches your job and run them on your own recorded calls, with your product names, your languages and the questions your customers really ask. Score each one on the job itself, such as whether a summary caught the callback time the caller was promised.

When a vendor pitches you, ask which model runs each step, from the live conversation to the routing and the summary after the call. Ask for the full name, with the generation number and the tier. One name for every step is worth a follow-up question, and our 30 questions to ask before buying voice AI cover what else to check.

Where does DRING come in?

The DRING stack is model-agnostic. The speech and reasoning models behind each AI agent are chosen for the language and the workflow, and each speech-to-text, language and text-to-speech provider is selected per agent from leading vendors. Any of them can be swapped without a rebuild and changed without touching the rest of the stack, as our page on the model layer explains.

A workflow can also route to a different model without a re-platform, and if a model provider has an issue, the call can fail over to a working one. Our page for IT directors goes through the questions an architecture review asks next.

A model change is tested like a launch. Each DRING agent runs 1,000 to 10,000 simulated conversations, built for your company, before the first real call. A stronger model goes live only after clearing the regression suite the whole stack runs, and results are compared against the previous release before anything ships. Tell us which step of your calls you want an AI model on first, and DRING calls you in two minutes.