KV-cache quantization without any fork (recommended, 2026): upstream llama.cpp/Ollama now cover this natively — use -ctk q8_0 -ctv q8_0 (half KV memory, negligible quality loss: perplexity +0.002–0.05) or -ctk q4_0 -ctv q4_0 (quarter memory, ≈7.6% perplexity increase). In Ollama: OLLAMA_KV_CACHE_TYPE=q8_0 with OLLAMA_FLASH_ATTENTION=1. Keep K and V types symmetric to stay on the fast fused Flash-Attention path. Since April 2026, mainline llama.cpp also applies Hadamard rotation to KV activations (PR #21038), which greatly improves low-bit KV quality (opt-out: LLAMA_ATTN_ROT_DISABLE=1).

The RotorQuant/TurboQuant fork flow below is experimental/legacy: the TurboQuant llama.cpp PR was closed without merging (June 2026) and the fork is unmaintained relative to mainline. It is NOT required to use this model.

Voxtral-Mini-3B-2507-RotorQuant-MLX-2bit

2-bit MLX weight-quantized build of mistralai/Voxtral-Mini-3B-2507 with a RotorQuant KV-cache profile. Ultra-compact, best-available 2-bit stability for streaming audio on Apple Silicon.

Hardware compatibility

Device	VRAM / RAM	Recommendation
Apple M4 Max 128 GB	~1.3 GB	recommended — headroom for long context
Apple M3 Max 64 GB	~1.3 GB	comfortable
Apple M2 Max 32 GB	~1.2 GB	fits

Overview

• Base: mistralai/Voxtral-Mini-3B-2507 — 3B speech-understanding model
• Capabilities: transcription, speech translation, audio QA
• Weight precision: 2-bit (group-wise)
• KV-cache profile: RotorQuant (rotational online re-basis)
• Approx. on-disk size: ~1 GB
• Runtime: MLX on Apple Silicon

RotorQuant's rotational re-basis helps 2-bit builds remain stable under distributional drift — preferred over TurboQuant at this precision for streaming workloads.

Quickstart

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("majentik/Voxtral-Mini-3B-2507-RotorQuant-MLX-2bit")

prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": [{"type": "audio", "path": "stream.wav"},
                                  {"type": "text", "text": "Transcribe."}]}],
    add_generation_prompt=True,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=256))

Model specs

Field	Value
Parameters	3B
Weight bits	2
Group size	32
Cache profile	RotorQuant
Size on disk	~1 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	Streaming, code-switching	Batch transcription

See also

• majentik/Voxtral-Mini-3B-2507-RotorQuant-MLX-4bit
• majentik/Voxtral-Mini-3B-2507-RotorQuant-MLX-8bit
• majentik/Voxtral-Mini-3B-2507-TurboQuant-MLX-2bit
• mistralai/Voxtral-Mini-3B-2507 — upstream base model

About the RotorQuant / TurboQuant labels

RotorQuant and TurboQuant are this project's release labels, not distinct quantization algorithms — for any given tier, both brand repos carry byte-identical weights produced with the standard MLX / llama.cpp quantizers. No brand-specific speedup is claimed or measured.