AI 101 — What are model parameters?

A parameter is a number inside an AI model that training adjusts, and the "70B" in a model's name means 70 billion of them. They are not facts, not rules and not code. They are the settings the model learned, and once training stops they are the model: copy those numbers into a file and you have shipped the whole thing.

Why the number sits in the name

Model names carry parameter counts because the count answers the question buyers actually ask: will this fit, and what will it cost to run? Llama 3 70B, Qwen 3 235B, an "8B" you can run on a laptop — the number is a size label, the way 1.5 litres is on a car engine.

Mixture-of-experts models have muddied that reading. Qwen3-235B-A22B holds 235 billion parameters in total but activates 22 billion for each token, according to its model card — 128 experts, eight switched on at a time. You need memory for the 235 billion; you pay compute closer to the 22 billion. That split is why two models with the same number on the label can behave completely differently in a bill.

The mental model

GPT-3 had 175 billion parameters, per the paper that introduced it. Nobody wrote them. Training began with random numbers and nudged every one of them, repeatedly, in whichever direction made the model better at predicting the next token in its training text. Do that across trillions of tokens and the numbers settle into a configuration that encodes grammar, style, reasoning habits and a blurry picture of the world.

So the model is a function: text in, a probability for every possible next token out, computed through dozens of layers of matrix multiplication. The parameters are the matrices. Nothing else is stored — there is no table of facts sitting behind the weights, which is exactly why models make things up and why retrieval systems exist.

The analogy: a mixing desk

Picture a recording studio's mixing desk, with a slider for every parameter. During training an infinitely patient engineer nudges each slider a hair at a time, checks whether the mix sounds better on the next example, and keeps going until the whole desk is dialled in across billions of songs.

Then the desk is locked and the sliders are welded in place. Using the model — inference — is pushing a signal through those locked settings: the audio comes out shaped by them, but nothing moves. This is the part people misread. Chatting with a model does not tune it; your conversation is a temporary note passed alongside the locked desk, not a change to it. Fine-tuning is the exception, and it means running the engineer again on your own data.

What the number actually predicts

Memory, mostly. Each parameter is stored at some precision: at 16 bits it takes two bytes, so 70 billion parameters is roughly 140 GB of weights before you add anything for the context being processed. At 4 bits per parameter the same model is about 35 GB — which is the whole reason quantization exists, and why people talk about a 70B model running on one consumer graphics card.

Cost per token, second. Every token generated sends an input through those matrices once, so a bigger model burns more compute per word unless an architecture like mixture of experts lets it skip most of itself.

Ability, only loosely. Parameter count correlates with capability because bigger models can hold more structure, and scaling laws describe how loss falls as models, data and compute grow. But the trend is not a promise. Training data quality, the number of tokens seen, and the post-training that turns a text predictor into an assistant all move the result at least as much. A well-trained 8B model from 2026 beats a carelessly trained 70B model from 2023 on most tasks you would care about.

Three things people get wrong

"More parameters means smarter." It means bigger. Capability depends on what the model was trained on and how, and distillation routinely produces small models that outperform larger ancestors on narrow work.

"Parameters are where the facts live." A single parameter does not mean "Paris is the capital of France". Knowledge is spread across millions of weights in a form that is hard to read and impossible to edit surgically — which is why correcting one wrong fact by retraining is expensive and unreliable.

"The model learns from me." Not from a chat, and not overnight. Weights change during training or fine-tuning, not during a conversation. What feels like memory is either text kept in the context window or a separate store the application writes to, a distinction that matters for privacy and for anyone promising a model that "gets to know you".

Where to learn more

The GPT-3 paper is the clearest look at what a parameter count meant at the start of this era, and modern model cards — Qwen's included — publish total and active counts side by side, which is now the honest way to read a name. For the practical side, the explainers on quantization and mixture of experts cover the two tricks that decide whether a given model fits on the hardware you have.

If a model's parameter count told you the whole story, what would you want the label to say instead? Write to us via our contact page.

AI 101 — What does GPT stand for?

GPT stands for Generative Pre-trained Transformer — the three-word recipe behind OpenAI's language models, from GPT-2 to today's GPT-6. It is also the most-abbreviated phrase in AI: most people who type "GPT" into a search box are really asking what the letters mean and why they keep appearing in the news. Why it matters right now Every model name in the headlines is a small riddle.

AI 101 — What are scaling laws?

Scaling laws are the empirical rules describing how an AI model's performance improves in a smooth, predictable curve as three things grow: the size of the model, the amount of data it trains on, and the computing power spent training it. They are not physics — nobody proved them in a lab the way gravity was proven — they are patterns that kept holding every time someone measured them, which is why the entire AI industry now plans around them. The pattern got its canonical statement in a…

AI 101 — What is AGI?

AGI — short for artificial general intelligence — is a hypothetical AI system that could do any intellectual job a person can do, instead of being good at one narrow task. Every AI you can actually use today is narrow: it writes, translates, spots patterns in scans, plays games — each system built for its lane. AGI names the destination where one system covers all the lanes.

AI 101 — What is a large language model?

A large language model (LLM) is a very big neural network trained on enormous amounts of text to do one job well: predict the next piece of a text, given everything that came before it — then tuned so that when you talk to it, it actually helps instead of rambling. ChatGPT, Claude, Gemini, Llama, DeepSeek: every chatbot in the news is an LLM, and most AI features in the software you already use have one quietly bolted inside. That prediction job sounds too simple to matter, and it is the whole…

AI 101 — What is overfitting?

Overfitting is when a model learns its training data too well — it gets the examples it has already seen almost perfect, then performs worse on anything new. Since the entire job of a model is to handle cases nobody showed it, overfitting is the most common way an AI system looks brilliant in a demo and disappoints in production. The term shows up in almost every model release, fine-tuning guide and benchmark write-up, and it is rarely explained.

AI 101 — What is reinforcement learning?

Reinforcement learning is how a machine learns by trial and error: it tries things, notices which attempts earned a reward, and drifts toward the behaviour that earns more of it. There is no answer key involved. In the supervised learning most people picture, someone hands the model millions of labelled examples and it learns to copy them. In reinforcement learning, nobody says what the right move is — the machine acts, gets a number back telling it how well that went, and adjusts.

AI 101 — What is LLM-as-a-judge?

If you read a week of AI news, you'll notice almost nobody grades models by hand any more. When a lab says its new model is better at writing, summarising or handling customers, the number usually came from another AI. An LLM judge is a language model used to grade the work of another model — or of your own AI feature — when no piece of code can decide whether the answer was good.

Google's Gemini 3.8 TTS clones a voice from 30 seconds of audio

Google put two text-to-speech models into the Gemini API today and made voice cloning a standard developer feature. Also: Meta's Connect keynote landed three products, and Microsoft's new Gulf number is mostly an old one. Google released Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS on Wednesday, and the headline capability is replication — building a synthetic voice from a reference recording as short as ten seconds, with Google's own marketing using thirty as the round number.