ggml-tiny.bin, the 77 MB whisper model: where it lives and when not to use it

If you typed the path to ggml-tiny.bin into a search box, you want one of two things: the exact file, or a straight answer on whether the smallest Whisper model is good enough for what you are building. This page gives you both. The download is one line. The second half is the part nobody writing about this file will tell you, because I build a voice-first Mac agent and we chose not to ship this model.

https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-tiny.bin

# from inside a cloned whisper.cpp checkout
./models/download-ggml-model.sh tiny
# -> writes models/ggml-tiny.bin (77.7 MB)

What ggml-tiny.bin actually is

GGML is the on-disk tensor format used by whisper.cpp, the C/C++ port of OpenAI's Whisper. The ggerganov/whisper.cpp repository on Hugging Face is a model store: it holds pre-converted GGML binaries so you do not need Python, PyTorch, or any conversion step. You download ggml-tiny.bin, point the whisper.cpp binary at it, and it runs.

"tiny" is the smallest of the five standard Whisper checkpoints, about 39 million parameters. That is what buys you the 77.7 MB file and the fast decode, and it is also what costs you accuracy. The whisper.cpp models README ranks it bluntly as the fastest and least accurate option. Everything else is a tradeoff up from there.

The model sizes, with exact byte counts

All of these live in the same repository. The byte counts below were read straight from the Hugging Face x-linked-size header on 2026-06-17, so you can sanity-check a download against them.

Quantized variants (the q5_1, q8_0 files) exist in the same repo and trade a little accuracy for a smaller footprint. Check the Files and versions tab on the repo for the current list before you hardcode a filename.

When tiny is genuinely the right call

I do not want to talk you out of this model. There are real jobs where 77.7 MB and a fast decode beat everything heavier:

• Wake-word and gating. You only need to know roughly what was said to decide whether to hand off to a bigger model. tiny is perfect as the cheap first pass.
• Constrained hardware. Old laptops, Raspberry Pi, anything where loading a 500 MB model is a non-starter. tiny runs where small will not.
• Strictly-offline requirements. If audio is not allowed to leave the machine, full stop, then a local model is the only honest answer and tiny is the lightest one.
• Batch transcription of clean audio. When latency does not matter and the audio is clean, tiny's mistakes are easy to skim past in a draft you are going to edit anyway.

Why Fazm does not ship ggml-tiny.bin

Fazm is a voice-first agent for macOS: you hold a hotkey, talk, and the same Claude Code / Codex agent loop acts on your machine. Voice is the front door, so transcription is on the hot path. You would assume we run a local Whisper model. We tried; we did not ship it.

The problem with tiny for a live agent is not mainly word accuracy, it is when the words arrive. whisper.cpp decodes a window of audio after that window closes. You finish a phrase, then you wait for the model to run, then you get text. For a dictation app that is fine. For an agent that should start reacting while you are still mid-sentence, that pause is the whole experience. The model that nails accuracy still loses if the user is staring at a frozen cursor.

So Fazm streams instead. Our TranscriptionService.swift hardcodes private let model = "nova-3" and opens a WebSocket to wss://api.deepgram.com/v1/listen. Audio goes up in chunks; interim (non-final) transcript segments come back mid-utterance, so the loop can move before you stop talking. We also feed a keyterm vocabulary list so command words and app names land, and we added a detector for the degenerate repeated-token loops that small ASR models emit on silence. That detector exists precisely because we saw how badly the lowest tier handles low-energy audio.

That is a real tradeoff, and I will state it plainly: streaming to Deepgram means audio leaves the machine. If your requirement is that it must not, ggml-tiny.bin (or a larger local Whisper model) is the correct choice and this is the page that tells you where to get it. We optimized for latency in the agent loop; you may weight privacy higher. Both are defensible.

If you landed here because you are wiring voice into something on a Mac and you are weighing local Whisper against a hosted streaming model, that is exactly the call Fazm made in production. You can read the agent it feeds into, or talk it through.