What total and active parameters tell you about a model

Each row in the tryingopen model picker reads like "Z.ai · 320B · Everyday questions, code and pictures". The middle number is the model's size in parameters, the learned numbers a model is made of.

That one number hides a split. Many models in the picker are mixtures of experts, which use only part of themselves for each token (a word or a piece of one). Makers give two numbers for such a model, and each sets a different cost.

What the picker's size shows

For almost every row (Gemma 4 E2B aside, covered below), the size is the total parameter count. Two names drift from their sizes: GPT-OSS 120B shows 117B and Gemma 4 26B shows 25B, following the makers' cards.

As of October 2026, three of the 24 models you can pick show no size:

• Qwen3.8 Flash is Alibaba's hosted production model, and we haven't found a size Alibaba gives for it. The 125B figure you may see belongs to Qwen3.8-Flash-Next, the open-weight preview it is based on.
• Mistral Large 4: Mistral's announcement describes "a 1 trillion-parameter natively multimodal model with 52 billion active parameters". Mistral says the weights will follow by the end of October.
• MiMo V2.6 Pro: OpenRouter's listing says "over 1T parameters". The card for Xiaomi's Pro-RL weights gives 1.02T total and 42B activated.

Total and active

In a dense model, every parameter works on every token. Gemma 4 31B (30.7B on Google's card), Qwen3.8 27B (per Alibaba's card) and Muse Glimmer 30B (about 29.6B, per Meta's card) are dense. Their total and active counts are the same.

A mixture of experts splits much of the model into blocks called experts. A small router inside the model picks a few experts for each token, and the rest sit idle. Tencent's Hy4 Preview has 256 routed experts plus 1 shared one, and each token goes through 8 of the routed ones.

LFM2.5 24B is a mixture of experts too: 24B total, with about 2B active by Liquid's blog or 2.45B by its card. Gemma 4 E2B shows Google's "effective" count, 2.3B, or 5.1B with its per-layer embeddings.

Total size sets the memory

Every expert has to sit in memory, because the router may pick any of them for the next token. So the total decides how much memory a model needs.

The rough arithmetic: at 16-bit precision each parameter takes 2 bytes, so a billion parameters needs about 2 GB. At 4-bit it needs about half a gigabyte. Working memory for the conversation comes on top. The makers' figures fit:

• OpenAI says gpt-oss-120b fits on a single 80 GB GPU because its expert weights are stored in MXFP4, a 4-bit format. All 117B still have to fit.
• Meta says a 30B model "at full precision" would need "over 55 GB", and about 4-bit precision brings Muse Glimmer's language model "under 20 GB".
• Liquid AI says LFM2.5-VL 3B runs in about 3 GB of memory.

Offline Base's edge devices are NVIDIA Jetson Orin Nano boards with 8 GB of memory shared by the CPU and GPU, so the models we run on them are stored at 4-bit. The boards run LFM2.5-VL 3B as a 1.67 GB 4-bit file from Liquid's GGUF repository, plus a vision encoder file. By the same arithmetic, GLM 5.3 Flash's weights need about 160 GB at 4-bit.

Active size sets the work per token

Each token passes through the active parameters only, so the active count sets the arithmetic per token. GPT-OSS 120B needs memory for 117B and does roughly the work of a 5B model per token.

Compare Nemotron 3.5 Lightning and Muse Glimmer 30B. Both are about 30B in total, so they need about the same memory. Lightning runs 3B per token and Glimmer runs all of its roughly 30B, about ten times the work.

The picker's groups go by total, so GPT-OSS 120B sits under "Big and powerful" while using 5.1B per token. The blurb in each row's "Learn more" panel often adds the active share. GLM 5.3 Flash's reads: "Only 18B of its 320B works on each word, so it is quick and cheap for its size." The speed you get also depends on the provider, and each reply shows it.

Where you see it on tryingopen

When you pick a cloud model, our small model also answers a chat's first two questions, and the side-by-side panel draws the two models as squares sized by parameter count. Its caption warns that many large models use only a slice of theirs for each word. Pick GLM 5.3 Flash and the panel says it has about 107 times the parameters.

The default pick shows the same split. The d1 router, Liquid AI's d1-3B decision model on our Jetsons, sends anything beyond a simple question from the 3B small model to GLM 5.3 Flash unless you pick another fallback (how d1 routes). That fallback uses 18B per token, six times the small model's whole 3B.

Why sources disagree on a size

A size can be the maker's headline total, one part of the model counted on its own, or every parameter in the published weight files. Hugging Face shows the last as "Model size" on a model's page.

DeepSeek V4.1 Flash: 763B or 552B

The picker shows 763B, Hugging Face's count of the published files. DeepSeek's card describes "552B backbone parameters" plus 196B of Engram memory, "sparsely accessed via token-based lookup", and gives no single total. Those add up to 748B, and the card doesn't explain the other 15B. It also gives two active counts: 8B per token reading your input and 16B writing the reply.

DeepSeek V4 Flash: 284B or 304B

The picker shows 284B and its blurb 13B active, as OpenRouter lists them. DeepSeek's V4.1 Flash card gives the V4 Flash base the same 284B backbone and 13B activated. The Hub metadata for the V4 Flash weights counts 304B, and the card doesn't break down the extra 20B.

Active counts move too. Mistral gives Large 4 52B active, and its Hugging Face placeholder says 49B per token, leaving out embeddings and output layers.

Reading a row

Read the number in the picker as a guide to memory. When a blurb or card names a smaller share per token, that share is the better guide to the work. When two sources disagree, check which count each one quotes. A tour of the 24 models in the picker explains the groups, bands and badges on each row.