Voxtral-Mini-4B-Realtime-2602: Self-Hosting & Deployment Guide
Voxtral-Mini-4B-Realtime-2602 is an open-weight speech recognition model that transcribes audio to text.
What is Voxtral-Mini-4B-Realtime-2602?
Voxtral-Mini-4B-Realtime-2602 is an open-weight speech recognition model that transcribes audio to text, published by Mistral AI with open weights at mistralai/Voxtral-Mini-4B-Realtime-2602. LLM API does not route this model today, so there is no endpoint or code sample for it here. What this page does give you is the self-hosting picture: published weight sizes, the VRAM each quantisation needs and the commands to serve it yourself. Every figure is read from the public repository and refreshed weekly. At import it had 1,748,846 downloads on Hugging Face in the previous 30 days, under the apache-2.0 licence.
How to run Voxtral-Mini-4B-Realtime-2602 yourself
Open weights (apache-2.0), 4.4B parameters. Below: how much memory the weights need at each quantization, which hardware fits them, and the commands to serve it.
Voxtral-Mini-4B-Realtime-2602 VRAM requirements
| Precision / quantization | Weights | Figure | Source |
|---|---|---|---|
| BF16 (full precision) | 8.9 GB | Estimate | 4.4B params × 2 bytes |
| FP8 | 5.8 GB | File size | ghecko78/Voxtral-Mini-4B-Realtime-2602-FP8-Dynamic |
| MLX 4-bit | 3.1 GB | File size | mlx-community/Voxtral-Mini-4B-Realtime-2602-4bit |
| GGUF Q8_0 | 4.7 GB | File size | handy-computer/Voxtral-Mini-4B-Realtime-2602-gguf |
| GGUF Q4_K_M | 2.8 GB | File size | handy-computer/Voxtral-Mini-4B-Realtime-2602-gguf |
File size = actual download size on Hugging Face. Estimate = parameter count × bytes per weight. Weights only — the KV cache for long contexts needs extra memory; the fit tables below add 15% headroom as a rule of thumb.
GPU requirements: deploy Voxtral-Mini-4B-Realtime-2602 on your hardware
| Hardware | BF16 (full precision) | FP8 | MLX 4-bit | GGUF Q8_0 | GGUF Q4_K_M |
|---|---|---|---|---|---|
| NVIDIA H200141 GB | 1× | 1× | 1× | 1× | 1× |
| NVIDIA H10080 GB | 1× | 1× | 1× | 1× | 1× |
| NVIDIA A100 80GB80 GB | 1× | 1× | 1× | 1× | 1× |
| NVIDIA L40S48 GB | 1× | 1× | 1× | 1× | 1× |
GPUs needed per variant.
| Hardware | GGUF Q8_0 | GGUF Q4_K_M |
|---|---|---|
| RTX 509032 GB | 1× | 1× |
| RTX 409024 GB | 1× | 1× |
| RTX 309024 GB | 1× | 1× |
GGUF via llama.cpp; more cards or CPU offload for larger files.
| Hardware | GGUF Q8_0 | GGUF Q4_K_M | MLX 4-bit |
|---|---|---|---|
| Mac, 512 GBM3 Ultra | Fits | Fits | Fits |
| Mac, 192 GBM2 Ultra | Fits | Fits | Fits |
| Mac, 128 GBM4 Max / M3 Max | Fits | Fits | Fits |
| Mac, 64 GBM4 Pro / Max | Fits | Fits | Fits |
Assumes ~75% of unified memory is usable by the GPU.
| Hardware | BF16 (full precision) | FP8 | MLX 4-bit | GGUF Q8_0 | GGUF Q4_K_M |
|---|---|---|---|---|---|
| AWS p5.48xlarge8× H100 · 640 GB | Fits | Fits | Fits | Fits | Fits |
| AWS p4de.24xlarge8× A100 80GB · 640 GB | Fits | Fits | Fits | Fits | Fits |
| AWS g6e.12xlarge4× L40S · 192 GB | Fits | Fits | Fits | Fits | Fits |
| Google Cloud a3-highgpu-8g8× H100 · 640 GB | Fits | Fits | Fits | Fits | Fits |
| Azure ND H100 v58× H100 · 640 GB | Fits | Fits | Fits | Fits | Fits |
Whole-instance GPU memory; public instance specs.
Not on LLM.API yet
We do not route this model through our API at the moment. Browse the models you can call today — most workloads have a close match already live.
Not available on LLM API yet
We do not route this model through our API at the moment, so there is no endpoint or code snippet for it yet. Browse the models you can call today — most workloads have a close match already live.
Browse available modelsWhen to Use — When NOT to Use
Use it if...
- You need transcripts of recordings or calls produced on your own hardware (Voxtral-Mini-4B-Realtime-2602)
- Audio must stay inside your network for privacy or compliance reasons
- The weights are published openly, so you can fine-tune on your own voices or vocabulary
- You are comparing self-hosting cost against a per-minute or per-character API
Avoid if...
- LLM API does not route this model today, so there is no endpoint here to call
- You want zero operations — serving, scaling and upgrades are yours to run
- You need a chat or reasoning model — this is an audio model
- You need a guaranteed real-time SLA before testing on your own audio
Get one key to every model
Swap your API key. Keep your code.