ggml-base.bin: the file, the size, and the one thing base gets wrong

You searched the exact /resolve/main/ path for ggml-base.bin, so you want one of two things: the actual file, or a straight answer on whether base is the model to build on. This page gives you both. The download is one line. The second half is the part the download guides skip, because I build a voice-first Mac agent and base's real limit is not the one its file size suggests.

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

# from inside a cloned whisper.cpp checkout
./models/download-ggml-model.sh base
# -> writes models/ggml-base.bin (~148 MB)

# or fetch it directly (follow the 302 redirect):
curl -L -o ggml-base.bin \
  https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.bin

base in context, with exact byte counts

Everything below lives in the same whisper.cpp model store. The byte counts were read straight from the Hugging Face x-linked-size header on 2026-06-18, so you can sanity-check a download against them. base is the row most projects quietly standardize on.

Quantized base variants (the q5_1 and q8_0 files) exist in the same repo. They shrink the file, not the accuracy ceiling. Check the Files and versions tab before you hardcode a filename, and read the official models README for the canonical size ranking.

What ggml-base.bin actually is, and why it is the default

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

base is the second of the five standard checkpoints, about 74 million parameters. People gravitate to it for a simple reason: it is the smallest model that stops obviously mangling common words, while still decoding faster than real time on Apple Silicon. tiny is cheaper but error-prone; small is more accurate but four times the size. base is the comfortable middle, which is exactly why so many tutorials default to it without saying why.

The thing base gets wrong: it writes "dot com", not ".com"

Here is the part nobody downloading this file mentions. base, like every vanilla Whisper checkpoint, was trained to transcribe natural speech. Say "open fazm dot ai slash download" and base gives you exactly that string of words. It has no concept that you meant a URL. Dictate a file path, a shell command, or an email address and you get the spoken form, not the written one. Going from base to small does not fix this; a bigger model spells the words better, it still does not know they were a URL.

This is the layer that matters once you try to turn any Whisper model into a real voice interface, and it sits entirely on top of the model. In Fazm, that layer is a defaultReplacements table in Desktop/Sources/TranscriptionService.swift: a list of spoken-form to written-form rules applied to the raw transcript. A sample:

The point is not that Fazm's table is special. It is that any serious use of base for command or code dictation needs something like it, and reaching for a heavier Whisper model will not give it to you. If you are wiring base into a voice flow, budget for the rewrite layer up front. It is cheaper than the model bump you would otherwise reach for, and it fixes a class of errors model size never touches.

Choosing base, honestly

base is a good default. Here is where it lands relative to the models on either side, and the one place it stops being a model-size question at all.

Why Fazm does not decode base.bin locally

Fazm is a voice-first agent for macOS: hold a hotkey, talk, and the same Claude Code or Codex agent loop acts on your machine. Voice is the front door, so transcription is on the hot path. You would expect us to ship a local Whisper model like base. We chose not to, for a reason orthogonal to accuracy.

whisper.cpp decodes a window of audio after that window closes. You finish a phrase, the model runs, then text appears. For dictation that pause is fine. For an agent that should start reacting while you are still mid-sentence, the pause is the whole experience. So Fazm streams instead: TranscriptionService.swift opens a WebSocket to a real-time ASR (Deepgram Nova-3) and gets interim transcript segments back mid-utterance, then applies the same defaultReplacements rewrite layer described above. The rewrite layer is the same problem whether you run base locally or stream; the streaming choice is purely about when the words arrive.

That is a real tradeoff, stated plainly: streaming means audio leaves the machine. If your requirement is that it must not, then ggml-base.bin (or a larger local Whisper model) is the correct choice, and this is the page that tells you where to get it and what to bolt on so it reads back ".com" instead of "dot com". We optimized for latency in the agent loop; you may weight privacy higher. Both are defensible.