inferoid-notes-e4b

The notes model behind Inferoid Notes: meeting notes written on your Mac, from a transcript of the call. Nothing leaves the Mac.

  • Base: Gemma 4 E4B, fine-tuned for meeting notes, MLX 4-bit (3.9 GB).
  • For Macs with 16 GB of memory or more.
  • Output: JSON with title, summary, key points, decisions, action items (owner, due date) and open questions.

Results

On 23 held-out synthetic business calls (written transcripts with speech-recognition-style errors, roughly 35-70 minutes of speech each), judged by independent models:

model coverage mistakes per call usefulness (1-10) median time per call (24 GB M4)
inferoid-notes-e4b 96.7% / 97.1% 3.0 / 2.9 7.9 / 7.3 40 s
Claude Opus 5.5 (cloud) 98.9% / 98.6% 3.2 / 3.0 8.8 / 8.0 cloud
Gemma 4 E4B, not fine-tuned (22 calls) 89.6% / 89.4% 1.3 / 1.1 8.2 / 7.4 37 s

Each cell: judge GPT-6.1 Sol / judge Grok 4.6. The fine-tuned model captures far more (coverage) at about the frontier model's mistake rate; the base model makes fewer mistakes because it writes much less.

Mistakes = statements in the notes that the call does not support (wrong owner, invented question, hedge stated as fact). These are synthetic calls, not recordings of real meetings.

Use

The easiest way is the app: uv tool install git+https://github.com/inferoid/inferoid then inferoid add notes.

Directly, with mlx-lm:

pip install mlx-lm
mlx_lm.server --model Inferoid/inferoid-notes-e4b --port 8090

Send the raw Gemma 4 prompt without the thinking token (<bos><|turn>system\n...<turn|>\n<|turn>user\n...<turn|>\n<|turn>model\n); the system prompt the app uses is in Engine.swift.

Downloads last month
86
Safetensors
Model size
7B params
Tensor type
U32
Β·
BF16
Β·
MLX
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Inferoid/inferoid-notes-e4b

Quantized
(463)
this model