LLMScalelive Hugging Face data

LLM model name decoder

What does each part of a model name mean? Type or pick a model — every token (size, quantization, variant…) is colour-coded and explained. Hover to highlight.

1

Model name

Breakdown
/---

Hover a token to light up its meaning.

RedHatAIPublisher

The org or user that published this repo (the model author or a quantizer).

Qwen3.6Model family

The base model line — often with its version number attached.

35BParameter count

Total parameters (billions). Drives quality and VRAM — bigger = more capable but heavier.

A3BMoE — active params

Active parameters per token (e.g. A3B = 3B active). Compute ≈ active params; VRAM ≈ total params.

NVFP4Quantization · NVFP4 / MXFP4

4-bit float (NVFP4 on Blackwell, MXFP4 e.g. in gpt-oss). ~4× smaller with better accuracy than INT4.

How a model name is read

An open-weight model's name packs several independent decisions into one string: the family and version, the size, the architecture, what it was tuned for, and how it was quantized or packaged. The decoder takes the organisation from before the slash, splits the rest on hyphens and underscores, and matches each part against rules for those categories, in order.

It recognises parameter counts (8B, 350M), Mixture-of-Experts sizes (8x7B) and active parameters (A3B); quantization and formats (AWQ, GPTQ, GGUF and its Q levels, FP8, NVFP4 and MXFP4, BF16, W4A16, INT4, bitsandbytes 4-bit); tuning (Instruct, Chat, Base, Coder, Math, VL, Embedding, Reranker, and reasoning names such as R1, QwQ and Thinking); training methods (DPO, ORPO, GRPO, LoRA); context lengths (32k), dates and versions. It reads the name only — conventions differ between families, and a name can be wrong about its own contents.

Questions

What does A3B mean in a model name?
About 3 billion active parameters per token in a Mixture-of-Experts model. Qwen3-30B-A3B holds 30 billion parameters but sends each token through experts totalling around 3 billion: memory is needed for all 30B, while the compute per token is closer to a 3B model's.
What is the difference between GGUF, AWQ and GPTQ?
GGUF is a file format for llama.cpp and the tools built on it, usually carrying quantization levels such as Q4 or Q8 for CPUs, Apple Silicon and consumer GPUs. AWQ and GPTQ are 4-bit weight quantization methods stored as safetensors and aimed at GPU serving engines such as vLLM.
What do FP8, NVFP4 and MXFP4 mean?
Floating-point weight formats smaller than BF16. FP8 stores each weight in one byte, half of BF16; NVFP4 and MXFP4 are 4-bit formats, half a byte per weight, used on recent NVIDIA hardware and by models such as gpt-oss.
What is the difference between Instruct and Base?
A base model is only pretrained to continue text. An Instruct model — also named Chat or -it — has been further tuned to follow instructions and hold a conversation, which is what most applications want.

Go deeper