liveBook · Manning

Parameter count

As explained in Build Applications with Local AI Models on a Mac

Model size is a measure of a language model's scale, commonly expressed by its number of learned parameters.

Parameter count describes a model’s size by the number of numerical values it learned during training. In model names, B means billions of parameters: gemma3:4b is a four-billion-parameter variant of Gemma 3. The count is a useful first clue when choosing a model, not a complete measure of what it can do. section 12.2 section 4.1.3

Within a model family, more parameters generally bring better responses, but also slower inference and greater memory needs. A smaller model may be the better choice when quick answers or limited laptop RAM matter most. Check available memory before downloading: actual requirements vary with quantization and model format. section 12.2 section 4.3.2

Table 4.1 Model sizes, approximate RAM requirements, and typical use cases

Model size

Approximate RAM needed

Typical use cases

1B–3B

2–4 GB

Quick Q&A, simple chat, text formatting

7B–9B

5–8 GB

General conversation, writing, coding help

13B–14B

10–16 GB

Complex analysis, detailed writing

27B–32B

20–32 GB

Advanced reasoning, professional tasks

70B+

40+ GB

Research, near-GPT-4 quality

Compare broad LLM size ranges with approximate RAM needs and typical tasks before choosing a download. section 4.3.2

Choosing a first local model

On an 8 GB laptop, gemma3:4b is a comfortable starting point; a 7B model may be tight, while a 70B model will not fit. Check current tags in the Ollama model library, because available models change. section 4.1.3 section 12.0

Table 9.2 MLX Whisper model sizes, memory requirements, and recommended use cases. The speeds were measured on a Mac Mini M2 for a 30-second English audio clip.

Model ID

Parameters

Download size

Memory (RAM)

30s audio speed

Accuracy

whisper-tiny-mlx

39 M

~75 MB

~200 MB

~1 s

Basic

whisper-base-mlx

74 M

~145 MB

~350 MB

~2 s

Good

whisper-small-mlx

244 M

~480 MB

~800 MB

~4 s

Very good

whisper-medium-mlx

769 M

~1.4 GB

~2 GB

~8 s

Excellent

whisper-large-v3-mlx

1550 M

~2.9 GB

~4 GB

~18 s

Near-human

Whisper provides a second comparison: its listed parameter counts rise alongside download size, memory use, transcription time, and accuracy. section 9.7

One caution: total parameters do not always describe how much of a model is active during a task. A mixture-of-experts model can have many total parameters but activate fewer at once. For a large download, check its model card or Ollama page for memory requirements rather than relying on the count alone. section 17.3.1

How Build Applications with Local AI Models on a Mac covers it

How deep it goes
Defines learned parameters and explains practical size tradeoffs, without detailing how those values are learned.
Examples
Compares gemma3:4b with larger laptop options and tabulates Whisper parameter counts against transcription speed and accuracy.
What it assumes
Introduces parameters from zero while assuming access to a machine whose available RAM can be checked.
What you can do after
Operate/configure it: choose an Ollama model using its size tag, available RAM, and current library listing.
How it connects
Connects parameter count to inference speed, memory requirements, quantization, and mixture-of-experts models.
Cover of Build Applications with Local AI Models on a Mac
liveBookmanning.com
Read Build Applications with Local AI Models on a Mac on liveBook Get the book at manning.com

Also explained in

Compare how 3 books explain parameter count →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Build Applications with Local AI Models on a Mac. The section links point to the parts of the book it draws on.