What does 8B mean in AI? 8 billion parameters, 5 GB to run

8B is a count of weights, and it predicts memory far better than quality

Three 8B models score 6, 7 and 11 on the Artificial Analysis Intelligence Index (a composite of public reasoning, coding and knowledge tests, read on 26 September 2026), and Kimi K3 stores 2.8 trillion parameters, 350 times as many, to score 44. The chart plots 13 open models with both numbers published; each count is the total from the vendor's own model card, and each score is the default mode except the three marked max effort, scored at maximum reasoning effort.

The dotted line is the frontier, the best score at each size, and it climbs 33 points across a 350 times increase in parameters. Below it, the spread at one size is as wide as the spread across sizes. Llama 4 Maverick stores 400 billion parameters and scores 10, one point under IBM's 8B.

What are parameters in AI? The learned numbers a model predicts with

A parameter is one learned number, a weight in the network that training adjusts and the model consults on every token it predicts. 8B counts 8 billion of them, and the B is billion the way the G in GB is giga. Meta's model card calls Llama 3.1 "a collection of pretrained and instruction tuned generative models in 8B, 70B and 405B sizes". Nothing else is being counted: not the training data, not the context window. What is an LLM covers how the weights get set.

Two more letters appear on model cards. M is million, where small embedding and vision models live, and T is trillion: Kimi K3's card says 2.8T. A name like Qwen3.6 35B A3B, or DeepSeek's "284B parameters, 13B activated", marks a mixture of experts model, which stores every parameter but uses only a slice (3 billion of 35, in the Qwen case) on any one token. Active parameters set the speed. Total parameters set the memory, and so decide whether the model loads at all.

How much memory does an 8B model need? 16 GB at 16 bit, about 5 GB quantized

An 8B model needs about 16 GB of memory at 16 bit and about 5 GB as the 4 bit download most people run, plus working memory for the conversation. The arithmetic is parameters times bytes per weight: 2 bytes in the 16 bit formats vendors publish (Hugging Face's memory guide counts an fp16 copy of the weights at 2 bytes per parameter), 1 byte at 8 bit, half a byte at 4 bit. Here is the Llama 3.1 family, which Ollama ships in all three sizes, checked on 26 September 2026.

The last column is our own division, download size in bits over the parameter count, and it shows the 4 bit rule of thumb undercounts what you actually download by about 20 percent: every Llama 3.1 size lands at 4.8 to 4.9 bits per weight, not 4. So budget 5 GB for an 8B, 43 GB for a 70B (Ollama's llama3.3 page lists Llama 3.3 70B at the same 43 GB) and 243 GB for a 405B.

That is why 8B is the size everyone talks about: 4.9 GB fits a laptop with 16 GB of RAM or an 8 GB graphics card. Mistral's card for Ministral 3 8B says it fits in 12 GB of VRAM at FP8 and less after quantization; its card for Mistral Small 3.1, a 24B model, says a single RTX 4090 or a 32 GB MacBook once quantized. A 70B at 43 GB needs a 64 GB Mac or two 24 GB cards, and above that you are renting a server.

Is 8B or 70B better? A 2026 8B beats a 2024 70B

Age beats size: IBM's Granite 4.2 8B, released 25 August 2026, scores 11 on the Intelligence Index, while Meta's Llama 3.3 70B from December 2024 scores 8 and Llama 3 70B from April 2024 scores 5, an Artificial Analysis estimate. The 70B has nearly nine times the weights. The count records how big the network is and says nothing about the data or recipe it was trained with, which is where two years of progress went. Scores below are as of 26 September 2026; parameter counts are the vendors' own.

Inside one family and one generation, size still pays: Granite 4.2 3B, 8B and 30B score 9, 11 and 15. Across generations it stops paying, and in the mixture of experts rows total and active diverge by 18 to 33 times, so the headline number is not even the one doing the work on each token.

The same weights can land 14 points apart. Qwen3.8 27B scores 20 in its default mode and 34 with reasoning effort turned up to xhigh, because the index also measures how long the model thinks before answering. What is a reasoning model covers that dial.

Why the GPT, Claude and Gemini flagships publish no parameter count

The closed flagships publish no size because the count is a competitive detail, and for weights you cannot download it tells you nothing you can use. OpenAI's last published count for a frontier GPT was GPT-3 at 175 billion, in the 2020 paper that introduced it; the sizes it does publish are for its open weights, and the August 2025 gpt-oss model card gives gpt-oss-120b 116.8 billion total and 5.1 billion active parameters. Google gave sizes only for its on-device Gemini Nano, 1.8 billion and 3.25 billion parameters in the Gemini 1.0 report, and none for Pro or Ultra. As of September 2026, GPT-6 Astra and Claude Fable 5.1 both score 53 on the Artificial Analysis leaderboard with no size beside them, nine points above Kimi K3, the largest open model at 2.8 trillion.

Read the count when you buy hardware, the score when you buy answers

The parameter count is the first number to read if you run a model yourself, because it decides whether the weights load, and close to the last if you use a hosted chat product. A hosted provider already bought the hardware, so what you pay tracks the token price, not the size.

For private or offline use, a hosted product is the wrong tool: download the 4.9 GB Llama 3.1 8B through Ollama, or the same weights through LM Studio, and it costs nothing past the disk space. The open source LLM guide covers the licenses.