README.md

metadata

base_model: mistralai/Voxtral-Mini-4B-Realtime-2602
library_name: mlx
license: apache-2.0
pipeline_tag: automatic-speech-recognition
tags:
  - voxtral
  - audio
  - speech
  - speech-recognition
  - realtime
  - streaming
  - asr
  - mlx
  - rotorquant
  - quantization
  - 2-bit

Voxtral-Mini-4B-Realtime-2602-RotorQuant-MLX-2bit

2-bit MLX weight-quantized build of mistralai/Voxtral-Mini-4B-Realtime-2602 with RotorQuant KV-cache. Ultra-compact real-time ASR for memory-constrained Apple Silicon — best-available 2-bit stability on streaming audio.

Overview

• Base: mistralai/Voxtral-Mini-4B-Realtime-2602 — 4B real-time ASR model
• Weight precision: 2-bit (group-wise)
• KV-cache profile: RotorQuant
• Approx. on-disk size: ~1.2 GB
• Runtime: MLX on Apple Silicon

Quickstart

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("majentik/Voxtral-Mini-4B-Realtime-2602-RotorQuant-MLX-2bit")

for chunk in audio_stream():
    prompt = tokenizer.apply_chat_template(
        [{"role": "user", "content": [{"type": "audio", "path": chunk}]}],
        add_generation_prompt=True,
    )
    emit(generate(model, tokenizer, prompt=prompt, max_tokens=32))

Model specs

Field	Value
Parameters	4B
Weight bits	2
Group size	32
Cache profile	RotorQuant
Size on disk	~1.2 GB
Target hardware	Apple Silicon (M1/M2/M3/M4)
License	Apache 2.0

RotorQuant vs TurboQuant

	RotorQuant	TurboQuant
Strategy	Rotational online re-basis	Per-head static calibration
Memory reduction	~4x on KV-cache	~3.5x on KV-cache
Best for	Noisy/multi-speaker streams	Predictable domains, lowest p50 latency

See also

• majentik/Voxtral-Mini-4B-Realtime-2602-RotorQuant-MLX-4bit
• majentik/Voxtral-Mini-4B-Realtime-2602-RotorQuant-MLX-8bit
• majentik/Voxtral-Mini-4B-Realtime-2602-TurboQuant-MLX-2bit
• majentik/Voxtral-Mini-4B-Realtime-2602-RotorQuant — KV-cache-only bundle
• mistralai/Voxtral-Mini-4B-Realtime-2602 — upstream base model