Zum Inhalt springen
ModelAufgelistetExtern
Browse files

Voxtral 4B TTS 2603 TurboQuant MLX 2bit

> [!TIP] > KV-cache quantization without any fork (recommended, 2026): upstream > llama.cpp/Ollama now cover this natively — use -ctk q80 -ctv q80 > (half KV memory, negligible quality loss: perplexity +0.002–0.05)...

Quellehuggingface.co/majentik/Voxtral-4B-TTS-2603-TurboQuant-MLX-2bit/tree/e45dcd8a2b0daf8434fdc01690dcdf20ae13bd20
ArtOtherBasismodellmistralai/Voxtral-4B-TTS-2603VersionRevision e45dcd8BBewertung

Modellquelle

Art
Other
Dateigröße
421–797 MB
Basismodell
mistralai/Voxtral-4B-TTS-2603
Version
vRevision e45dcd8
Aktualisiert
26. Okt. 2026
Downloads
32
Quelle
Hugging Face

Quellenbeschreibung

[!TIP] 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-4B-TTS-2603-TurboQuant-MLX-2bit

2-bit MLX weight-quantized build of mistralai/Voxtral-4B-TTS-2603 with a TurboQuant KV-cache profile. Extreme-compression TTS variant for memory-constrained Apple Silicon.

Hardware compatibility

DeviceVRAM / RAMRecommendation
Apple M4 Max 128 GB~1.6 GBrecommended — headroom for long context
Apple M3 Max 64 GB~1.6 GBcomfortable
Apple M2 Max 32 GB~1.4 GBfits

Overview

  • Base: mistralai/Voxtral-4B-TTS-2603 — 4B multilingual TTS with zero-shot voice cloning
  • Weight precision: 2-bit (group-wise)
  • KV-cache profile: TurboQuant
  • Approx. on-disk size: ~1.2 GB
  • Runtime: MLX on Apple Silicon

Expect minor naturalness degradation vs 4-bit. Works best for a stable single-voice, single-language deployment. For multi-voice batches consider the RotorQuant 2-bit variant.

Quickstart

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("majentik/Voxtral-4B-TTS-2603-TurboQuant-MLX-2bit")

prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": [
        {"type": "audio", "path": "reference_voice.wav"},
        {"type": "text", "text": "Short prompt."},
    ]}],
    add_generation_prompt=True,
)
audio_tokens = generate(model, tokenizer, prompt=prompt, max_tokens=1024)

Model specs

FieldValue
Parameters4B
Weight bits2
Group size32
Cache profileTurboQuant
Languages9
Voice cloningZero-shot
Size on disk~1.2 GB
Target hardwareApple Silicon (M1/M2/M3/M4)
LicenseApache 2.0

RotorQuant vs TurboQuant

TurboQuantRotorQuant
StrategyPer-head static calibrationRotational online re-basis
Memory reduction~3.5x on KV-cache~4x on KV-cache
Best forSingle-voice sessionsMulti-voice / multi-language batches

See also

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.

Quellen

1 Quelle
Hugging FacePrimärVersion e45dcd8a2b0daf8434fdc01690dcdf20ae13bd20
huggingface.co/majentik/Voxtral-4B-TTS-2603-TurboQuant-MLX-2bit/tree/e45dcd8a2b0daf8434fdc01690dcdf20ae13bd20

Verifiziert 26. Okt.

Download files

Modellartefakte

2 Artefakte

model-00001-of-00002.safetensors

safetensors · 797 MB · SHA-256 1ed1b3c3629f…078e · Hugging Face

Download shard

model-00002-of-00002.safetensors

safetensors · 421 MB · SHA-256 cad90f782d17…8e1c · Hugging Face

Download shard

Quellenauszüge

3 Auszüge
Readme7. Aug.huggingface.co/majentik/Voxtral-4B-TTS-2603-TurboQuant-MLX-2bit

--- base_model: mistralai/Voxtral-4B-TTS-2603 library_name: mlx license: apache-2.0 pipeline_tag: text-to-speech tags: - voxtral - audio - speech - tts - text-to-speech - voice-cloning - zero-shot - mlx - turboquant - quantization - 2-bit --- > [!TIP] > **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](https://github.com/ggml-org/llama.cpp/pull/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. <!-- kv-upstream-note --> # Voxtral-4B-TTS-2603-TurboQuant-MLX-2bit 2-bit MLX weight-quantized build of [`mistralai/Voxtral-4B-TTS-2603`](https://huggingface.co/mistralai/Voxtral-4B-TTS-2603) with a TurboQuant KV-cache profile. Extreme-compression TTS variant for memory-constrained Apple Silicon. ## Hardware compatibility | Device | VRAM / RAM | Recommendation | | --- | --- | --- | | Apple M4 Max 128 GB | ~1.6 GB | recommended — headroom for long context | | Apple M3 Max 64 GB | ~1.6 GB | comfortable | | Apple M2 Max 32 GB | ~1.4 GB | fits | ## Overview - **Base:** `mistralai/Voxtral-4B-TTS-2603` — 4B multilingual TTS with zero-shot voice cloning - **Weight precision:** 2-bit (group-wise) - **KV-cache profile:** TurboQuant - **Approx. on-disk size:** ~1.2 GB - **Runtime:** MLX on Apple Silicon Expect minor naturalness degradation vs 4-bit. Works best for a stable single-voice, single-language deployment. For multi-voice batches consider the RotorQuant 2-bit variant. ## Quickstart ```bash pip install mlx-lm ``` ```python from mlx_lm import load, generate model, tokenizer = load("majentik/Voxtral-4B-TTS-2603-TurboQuant-MLX-2bit") prompt = tokenizer.apply_chat_template( [{"role": "user", "...

Note26. Okt.huggingface.co/majentik/Voxtral-4B-TTS-2603-TurboQuant-MLX-2bit

Source context: 32 downloads · 0 likes · Pipeline text-to-speech · Library mlx · Repo majentik/Voxtral-4B-TTS-2603-TurboQuant-MLX-2bit

Note7. Aug.huggingface.co/majentik/Voxtral-4B-TTS-2603-TurboQuant-MLX-2bit

Source context: 116 downloads · 0 likes · Pipeline text-to-speech · Library mlx · Repo majentik/Voxtral-4B-TTS-2603-TurboQuant-MLX-2bit