Model names are not standardized, but the labs that publish weights mostly follow the same habits, and a name usually carries four things in a fixed order. Take Qwen3-235B-A22B, the largest model in the Qwen3 family as the Qwen team’s release post lists it. Qwen3 is the family and generation. 235B is the number of parameters, the learned numbers inside the model, written in billions. A22B says that 22 billion of them are activated for any one token, a token being the chunk of text, often a word or part of one, that the model reads or writes at each step. The model card gives both figures plainly: 235B total, with 22B activated. The two numbers differ because the model is a mixture of experts, and the difference between them is the first thing to understand about a modern name.
In a dense model every parameter takes part in every token. In a sparse mixture of experts, part of each layer is copied many times, the copies are called experts, and a small router network picks a few of them for each token. The Mixtral paper, from Mistral AI, gives the cleanest account. Each layer of Mixtral 8x7B has eight feedforward blocks, and the router selects two for every token at every layer, so a token can draw on 47 billion parameters but only 13 billion are active during inference. The paper calls the first figure the sparse parameter count and the second the active parameter count, and it ties the memory needed to serve the model to the sparse count and the compute per token to the active one. The “8x7B” in the name is not a multiplication to trust: eight experts of seven billion would be 56 billion, and the paper says 47, because what it repeats eight times is the feedforward block of each layer, not the whole network. Qwen3-235B-A22B has 128 experts and activates 8 per token, according to its card. OpenAI’s gpt-oss-120b lists 117 billion parameters with 5.1 billion active, and its smaller sibling 21 billion with 3.6 billion. So the first number in a name tells you roughly how much memory the weights need, and the A-number tells you how much arithmetic each token costs. Compare total against total for memory, and active against active for speed.
After the size come words describing what the weights were trained to do. Training happens in stages. Meta’s Llama 3 paper releases its flagship in “pre-trained” and “post-trained” versions, and the Qwen team does the same: next to the post-trained Qwen3-30B-A3B sits Qwen3-30B-A3B-Base, the pre-trained counterpart. A base model is the product of pre-training alone, and a post-trained model has been trained further, which is where an assistant’s habits come from. The InstructGPT paper describes one recipe for that step: fine-tune on demonstrations written by human labelers, then train a reward signal from human rankings of outputs and apply reinforcement learning from human feedback. Its starting point is the sentence that making language models bigger does not inherently make them better at following a user’s intent. A suffix such as Instruct is the usual label for a model that went through this stage, though each lab defines its own. Thinking marks a model that writes out its reasoning before it answers. The original Qwen3-235B-A22B card describes one model that does both, switched by a setting, while the Qwen3-235B-A22B-Instruct-2507 card says it supports only non-thinking mode and the Thinking-2507 card says it supports only thinking mode. Neither card says what 2507 means. Treat a suffix like that as a version label, and read the card for what it changed, because the name will not tell you.
A name may also carry a number format. Quantization stores a model’s weights in lower precision to cut the memory it needs, while trying to preserve accuracy, as the Hugging Face documentation puts it. Weights are typically stored in 32-bit or 16-bit floating point, and a quantized copy might use 8 or 4 bits for each. FP8 is an 8-bit floating-point format, defined in a paper with two variants, E4M3 and E5M2. AWQ, short for Activation-aware Weight Quantization, is a weight-only method that protects the roughly one percent of weights the authors found most important, and Hugging Face lists it as a 4-bit method. GGUF is a file format made by the developer of llama.cpp, which stores the tensors together with metadata and is read by llama.cpp, LM Studio, GPT4All and Ollama; Hugging Face’s table of its quantization types lists Q4_K at 4.5 bits per weight and Q6_K at about 6.6. The arithmetic is simple: a parameter at 16 bits takes two bytes, so 235 billion of them come to about 470 GB, at 8 bits about 235 GB, and at 4 bits about 117 GB, for the weights alone. That is why gpt-oss-120b, whose card says its mixture-of-experts weights are quantized with MXFP4, can run on a single 80 GB GPU, the memory NVIDIA lists for an H100. A quantized copy is a different file from the original, often made by someone other than the lab, and its quality is a claim to check, not a given.
Last, what a parameter count does not tell you. It sets the size of the file and, for a dense model, the cost of each token; it does not set the quality. The Chinchilla paper trained a 70-billion-parameter model on about four times more data than Gopher, which has 280 billion, with the same compute budget, and found it outperformed Gopher, GPT-3, Jurassic-1 and Megatron-Turing NLG across many tasks. The InstructGPT authors report that human raters preferred the outputs of their 1.3-billion-parameter model to those of the 175-billion-parameter GPT-3. The number of parameters says how much a model can hold. What it was trained on, for how long and with what finishing steps decides what it does with it, and for models whose makers publish little, the count itself may be an estimate: Epoch AI’s database labels parameter counts, as it does training compute, by how far they can be trusted. The next lesson turns to the arithmetic of training.