Skip to content
English

Switch theme

// Mistral / Magistral / Devstral · Speech-to-text

Voxtral-Mini-4B-Realtime-2602: VRAM & local deployment

4.43B parameters. About 3.5 GB at 4-bit and 10.2 GB at BF16.

released 2026-01-21licence: Commercial OKmemory figures are estimates

01Specs

PublisherMistral AI · mistralai
Family / seriesMistral / Magistral / Devstral · Voxtral-Mini-Realtime
TypeSpeech-to-text (automatic-speech-recognition)
Parameters4.43B Source: counted from official safetensors weights
Released weightsBF16 · 17.7 GBdownload
LicenceApache-2.0 · Commercial OKPermissive licence (Apache-2.0 / MIT / CC-BY / OpenMDW): commercial use allowed.
Released2026-01-21 (Hugging Face repo created)

02How much VRAM it needs (estimate)

Speech models are small; the weight size is close to the total memory needed.

PrecisionWeightsWeights + overheadSmallest GPUMac
BF16 / FP16 (original)8.9 GB 10.2 GB 12GB GPUMac 24GB
FP8 / INT8 / Q8_04.7 GB 5.7 GB 8GB GPUMac 16GB
4-bit (Q4_K_M / AWQ / GPTQ-Int4)2.7 GB 3.5 GB 8GB GPUMac 16GB

Bits per weight: BF16 16, FP8 8.5, 4-bit 4.85. “Fits” = total ≤ 90% of memory; Mac GPU can use ~2/3 of unified memory (≤36GB) or ~3/4 (larger). For MoE, all experts must be in memory, so size by total parameters. Method →

03How to run it

Only runtimes the publisher or the runtime’s official library names for this model.

Which inference engine to choose →

04Official repositories & quantised builds

1 official repo(s) from mistralai.

05Other Mistral / Magistral / Devstral models

All 35 →
ModelParams4-bit totalReleased
LIDstral-Arabic——2026-10-08
Voxtral-Mini-4B-Realtime-Arabic4.43B3.5 GB2026-10-08
Shieldstral-1.0-3B3.85B3.9 GB2026-07-16
Mistral-Medium-3.5-128B127.7B88.6 GB2026-03-31
Mistral-Small-4-119B-2603119B / 6.5B80.0 GB2026-01-23
Devstral-2-123B-Instruct-2512125B86.8 GB2025-11-28
Mistral-Large-3-675B-Instruct-2512675B / 41B451 GB2025-11-28
Voxtral-4B-TTS-26034B3.2 GB2025-11-17

06Sources

Data checked 2026-03-11. Parameters, context and licence are copied from the official repository; memory figures are our estimates.

Data updated: Self-host open models(UTC+8)