Voxtral-Mini-4B-Realtime-Arabic Speech Recognition Model: 4GB+ VRAM

Sample outputs are available on the model card.
- 1. At a Glance
- 2. Overview
- 3. Specifications
- 4. Performance and Quality
- 5. Strengths and Use Cases
- 6. Hardware Requirements
- 7. Can You Run It Locally?
- 8. Our Own Measurements
- 9. Recent Models in the Same Size Class
- 10. Distributed Files
- 11. How to Get It
- 12. Quantized and Converted Variants
- 13. Related Articles
- 14. What to Read Next
- 15. Sources
- 16. Update History
At a Glance
| Item | Value |
|---|---|
| Repository | mistralai/Voxtral-Mini-4B-Realtime-Arabic |
| Family guide | Ministral-3 guide (2 articles) |
| Publisher guide | Mistral AI: models and licenses |
| Published | 2026-10-08 |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Mistral AI has released “Voxtral-Mini-4B-Realtime-Arabic", a streaming speech recognition model supporting Arabic dialects and Modern Standard Arabic. It is a Speech-to-Text model that outputs text in real time as audio is input, designed for use cases such as live subtitles and voice interfaces.
This model is a fine-tuned variant of the base model “Voxtral-Mini-4B-Realtime-2602". It boasts strengths in handling code-switching—where speakers shift between multiple languages within the same conversation—and challenging real-world audio inputs. It is positioned as a sovereign AI asset co-developed under a strategic partnership with Morocco’s Ministry of Digital Transition and Administrative Reform (MTNRA).
Specifications
Specifications based on the published model card and related materials are as follows:
- Parameter count: Approximately 4.4 billion (BF16 precision)
- Architecture:
VoxtralRealtimeForConditionalGeneration, consisting of a causal audio encoder and a language decoder - Input audio specification: 16 kHz sampling rate
- Output specification: Sequential text generation as audio streams are received (streaming output)
- Recommended settings: Temperature
0.0, sampling disabled (do_sample=False) - Latency settings: Configurable transcription delay. Benchmark evaluations by the publisher set this to 480 ms (base model specifications in tekken.json allow selecting from 80 ms to 1200 ms in 80 ms increments, as well as 2400 ms)
- Supported languages and dialects: Modern Standard Arabic, Moroccan Darija, and various other Arabic dialects (tag information includes ar, ary, ayl, aeb, arq, mey, afb, ars, acx, ayn, arz, apd, acm, ajp, apc, shu, etc.)
Performance and Quality
According to benchmark results published by the publisher, Voxtral-Mini-4B-Realtime-Arabic achieved an average CER (Character Error Rate: transcription error rate by character, where lower is better) of 8.82% across seven Arabic benchmarks with the transcription delay set to 480 ms. This score approaches within 0.91 points of the CER of 7.91% recorded by the offline-only model “Voxtral Transcribe Arabic", and is reported to outperform multiple open-source and proprietary offline models.
In addition to existing Arabic benchmarks, two new benchmarks co-developed with MTNRA, “ISMA" and “Darija in the Wild", were used for evaluation. Normalization code (normalization.py) that handles Arabic dialect-specific spellings, clitics, loanwords, and spoken numerals was applied when calculating the CER during evaluation.
Additionally, the publication materials for the base model “Voxtral-Mini-4B-Realtime-2602" include the following comparison table on the multilingual speech benchmark “Fleurs" (showing performance figures published for the base model):
→ Scroll horizontally to see all columns
| Model | Delay | AVG | Arabic | German | English | Spanish | French | Hindi | Italian | Dutch | Portuguese | Chinese | Japanese | Korean | Russian |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Voxtral Mini Transcribe 2.0 | Offline | 5.90% | 13.54% | 3.54% | 3.32% | 2.63% | 4.32% | 10.33% | 2.17% | 4.78% | 3.56% | 7.30% | 4.14% | 12.29% | 4.75% |
| Voxtral Mini 4B Realtime 2602 | 480 ms | 8.72% | 22.53% | 6.19% | 4.90% | 3.31% | 6.42% | 12.88% | 3.27% | 7.07% | 5.03% | 10.45% | 9.59% | 15.74% | 6.02% |
| 160 ms | 12.60% | 24.33% | 9.50% | 6.46% | 5.34% | 9.75% | 15.28% | 5.59% | 11.39% | 10.01% | 17.67% | 19.17% | 19.81% | 9.53% | |
| 240 ms | 10.80% | 23.95% | 8.15% | 5.91% | 4.59% | 8.00% | 14.26% | 4.41% | 9.23% | 7.51% | 13.84% | 15.17% | 17.56% | 7.87% | |
| 960 ms | 7.70% | 20.32% | 4.87% | 4.34% | 2.98% | 5.68% | 11.82% | 2.46% | 6.76% | 4.57% | 8.99% | 6.80% | 14.90% | 5.56% | |
| 2400 ms | 6.73% | 14.71% | 4.15% | 4.05% | 2.71% | 5.23% | 10.73% | 2.37% | 5.91% | 3.93% | 8.48% | 5.50% | 14.30% | 5.41% |
Looking at the base model’s measurement results, longer delays reduced error rates, improving the Arabic CER to 14.71% at the 2400 ms setting, whereas it remained at 22.53% at the 480 ms setting. Furthermore, compared to English (4.90%) and Italian (3.27%), the base model alone exhibited higher error rates for Arabic. The specialized model released this time is configured to achieve significant recognition accuracy improvements despite low latency (480 ms) by intensively training on this Arabic domain.
Strengths and Use Cases
Voxtral-Mini-4B-Realtime-Arabic is a model specialized for Arabic transcription in streaming environments where low latency is required. Primary use cases include real-time live subtitle generation and voice interfaces such as voice dialogue systems.
A major strength of this model is its support for a diverse range of Arabic dialects in addition to Modern Standard Arabic. It is tuned to withstand audio environments containing regional dialects like Moroccan Darija as well as real-world specific noise and phrasing.
In addition, adaptability to “code-switching", where speakers switch between multiple languages mid-conversation, is emphasized. According to the model card, intensive training on Arabic-centric code-switching resulted in reported accuracy improvements not only in anticipated combinations but also in code-switching involving languages outside the training focus.
On the other hand, this model primarily targets Arabic speech recognition. As a limitation, using the base model “Voxtral-Mini-4B-Realtime-2602" is recommended when performing speech transcription targeting multilingual languages other than Arabic.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 4.4B parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 4GB (laptop iGPU / phone class) | Q6_K | 2.6GB | 3.2GB |
| 8GB (RTX 4060 / 3060 Ti, etc.) | F16 | 6.4GB | 7.7GB |
Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: not registered, MLX (mlx-lm): not registered. “Not registered" means the name is absent from that registry today, not that the model cannot run.
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build mradermacher/Voxtral-Mini-4B-Realtime-Arabic-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
The publisher distributes this model as safetensors.
Compression: the Q6_K build measures 5.04 bits per weight — about 32% the size of the original 16-bit weights, calculated by this site from the actual file sizes.
Compiled by this site’s code from the published formats and the license field.
Our Own Measurements
Measurements We Did Not Take
This model does not state a license, so we cannot confirm permission to use it. Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).
Recent Models in the Same Size Class
Models with 4–15B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.
→ Scroll horizontally to see all columns
| Model | Parameters | Smallest VRAM tier | License | Our article |
|---|---|---|---|---|
| ggml-org/VibeVoice-ASR-GGUF | 8.3B | 8GB | mit | VibeVoice-ASR-GGUF Speech Recognition and Diarization Model: 8GB+ VRAM (2026-11-17) |
| Edge0/Audio8-ASR-Infinite | 4.1B | 12GB | apache-2.0 | Audio8-ASR-Infinite Speech Recognition Model: 12GB+ VRAM, File List (2026-09-24) |
Distributed Files
Weight files published in mistralai/Voxtral-Mini-4B-Realtime-Arabic, listed by this site from the Hugging Face API. Sizes are the actual file sizes.
| File | Size |
|---|---|
consolidated.safetensors |
8.86GB |
model.safetensors |
8.86GB |
How to Get It
The weights for Voxtral-Mini-4B-Realtime-Arabic are published in safetensors format on the Hugging Face repository. The repository is not gated, allowing downloads without additional consent requests.
This model supports vLLM and Transformers as inference engines.
When using vLLM, you can install the vLLM package and mistral-common audio dependencies, then start the server with the following command:
pip install --upgrade vllm "mistral-common"
vllm serve mistralai/Voxtral-Mini-4B-Realtime-Arabic
After starting the server, you can use real-time transcription by streaming audio via WebSocket to the /v1/realtime endpoint.
When using Transformers, native support is available in version 5.2.0 or higher. Install the necessary dependencies with the following commands:
pip install --upgrade "transformers>=5.2.0" accelerate
pip install --upgrade "mistral-common"
An example code snippet for running it from a Python script is as follows. It loads the model and processor, passes an audio array resampled to 16 kHz, and performs inference:
import torch
from transformers import VoxtralRealtimeForConditionalGeneration, AutoProcessor
from mistral_common.tokens.tokenizers.audio import Audio
from huggingface_hub import hf_hub_download
repo_id = "mistralai/Voxtral-Mini-4B-Realtime-Arabic"
processor = AutoProcessor.from_pretrained(repo_id)
model = VoxtralRealtimeForConditionalGeneration.from_pretrained(
repo_id, dtype=torch.bfloat16, device_map="auto"
)
audio_file = hf_hub_download(
repo_id=repo_id,
filename="assets/bank-take-2.wav",
repo_type="model",
)
audio = Audio.from_file(audio_file, strict=False)
audio.resample(processor.feature_extractor.sampling_rate)
inputs = processor(audio.audio_array, return_tensors="pt")
inputs = inputs.to(model.device, dtype=model.dtype)
outputs = model.generate(**inputs, do_sample=False)
print(processor.batch_decode(outputs, skip_special_tokens=True)[0])
Quantized and Converted Variants
→ Scroll horizontally to see all columns
| Added | Publisher | Format | Repository | Smallest VRAM tier (build, est. memory) |
|---|---|---|---|---|
| 2026-11-17 | mradermacher | GGUF | mradermacher/Voxtral-Mini-4B-Realtime-Arabic-GGUF | Q6_K 3.2GB (fits in 4GB VRAM) |
File sizes of each build:
- Available builds in mradermacher/Voxtral-Mini-4B-Realtime-Arabic-GGUF: Q2_K 1.4GB / Q3_K_S 1.5GB / Q3_K_M 1.7GB / Q3_K_L 1.8GB / IQ4_XS 1.8GB / Q4_K_S 1.9GB / Q4_K_M 2.0GB / Q5_K_S 2.3GB / Q5_K_M 2.3GB / Q6_K 2.6GB / Q8_0 3.4GB / F16 6.4GB
In addition, 1 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.
This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.
Related Articles
- Voxtral-Mini-4B-Realtime-2602-GGUF Real-time ASR Model: 4GB+ VRAM
- ZoomGain-4B: 4GB+ VRAM, GGUF Builds
- Mistral Large 4 Released: Multimodal MoE Model
- Mistral AI Announces Voxtral 3 Audio Models and Voice Agent SDK
What to Read Next
- Find models by VRAM (This model runs from the 4GB tier) → Other models that run on a 8GB GPU
- Explore the same model family → Ministral-3 family overview (2 articles, 1 converted builds)
- What Q6_K mean and where to get this model → GGUF format guide and models / Quantization and model-format glossary
- Learn about the publisher → Mistral AI: models, licenses and articles
- Other models for the same task → Other audio models
Sources
Update History
- 2026-11-17: Added converted builds to “Quantized and Converted Variants”: mradermacher/Voxtral-Mini-4B-Realtime-Arabic-GGUF
- 2026-11-17: Updated the hardware requirements table with the actual file sizes of mradermacher/Voxtral-Mini-4B-Realtime-Arabic-GGUF.
- 2026-11-17: Changed the title to show what the article covers (VRAM requirements, file list, etc.).

