Model Overview
- Model Architecture: VoxtralForConditionalGeneration
- Input: Audio-Text
- Output: Text
- Model Optimizations:
- Weight quantization: FP8
- Activation quantization: FP8
- Intended Use Cases: Voxtral builds upon Ministral-3B with powerful audio understanding capabilities.
- Dedicated transcription mode: Voxtral can operate in a pure speech transcription mode to maximize performance. By default, Voxtral automatically predicts the source audio language and transcribes the text accordingly
- Long-form context: With a 32k token context length, Voxtral handles audios up to 30 minutes for transcription, or 40 minutes for understanding
- Built-in Q&A and summarization: Supports asking questions directly through audio. Analyze audio and generate structured summaries without the need for separate ASR and language models
- Natively multilingual: Automatic language detection and state-of-the-art performance in the world’s most widely used languages (English, Spanish, French, Portuguese, Hindi, German, Dutch, Italian)
- Function-calling straight from voice: Enables direct triggering of backend functions, workflows, or API calls based on spoken user intents
- Highly capable at text: Retains the text understanding capabilities of its language model backbone, Ministral-3B
- Release Date: 08/21/2025
- Version: 1.0
- Model Developers: Mistral
- ModelCar Storage URI: oci://registry.stage.redhat.io/rhelai1/modelcar-voxtral-mini-3b-2507-fp8-dynamic:1.5
- Validated on RHOAI 2.25: quay.io/modh/vllm:rhoai-2.25-cuda
- Validated on RHAIIS 3.2.2: http://registry.redhat.io/rhaiis/vllm-cuda-rhel9:3.2.2
- Validated on vLLM: 0.10.1.1
Quantized version of Voxtral-Mini-3B-2507.
Model Optimizations
This model was obtained by quantizing activation and weights of Voxtral-Mini-3B-2507 to FP8 data type. This optimization reduces the number of bits used to represent weights and activations from 16 to 8, reducing GPU memory requirements (by approximately 50%) and increasing matrix-multiply compute throughput (by approximately 2x). Weight quantization also reduces disk size requirements by approximately 50%.
Only weights and activations of the linear operators within transformers blocks of the language model are quantized. Weights are quantized with a symmetric static per-channel scheme, whereas activations are quantized with a symmetric dynamic per-token scheme. The llm-compressor library is used for quantization.
Deployment
Use with vLLM
- Initialize vLLM server:
vllm serve RedHatAI/Voxtral-Mini-3B-2507-FP8-dynamic --tokenizer_mode mistral --config_format mistral --load_format mistral
- Send requests to the server, according to the use case. See the following examples.
from mistral_common.protocol.instruct.messages import TextChunk, AudioChunk, UserMessage, AssistantMessage, RawAudio
from mistral_common.audio import Audio
from huggingface_hub import hf_hub_download
from openai import OpenAI
# Modify OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://<your-server-host>:8000/v1"
client = OpenAI(
api_key=openai_api_key,
base_url=openai_api_base,
)
models = client.models.list()
model = models.data[0].id
obama_file = hf_hub_download("patrickvonplaten/audio_samples", "obama.mp3", repo_type="dataset")
bcn_file = hf_hub_download("patrickvonplaten/audio_samples", "bcn_weather.mp3", repo_type="dataset")
def file_to_chunk(file: str) -> AudioChunk:
audio = Audio.from_file(file, strict=False)
return AudioChunk.from_audio(audio)
text_chunk = TextChunk(text="Which speaker is more inspiring? Why? How are they different from each other?")
user_msg = UserMessage(content=[file_to_chunk(obama_file), file_to_chunk(bcn_file), text_chunk]).to_openai()
print(30 * "=" + "USER 1" + 30 * "=")
print(text_chunk.text)
print("\n\n")
response = client.chat.completions.create(
model=model,
messages=[user_msg],
temperature=0.2,
top_p=0.95,
)
content = response.choices[0].message.content
print(30 * "=" + "BOT 1" + 30 * "=")
print(content)
print("\n\n")
# The speaker who is more inspiring is the one who delivered the farewell address, as they express
# gratitude, optimism, and a strong commitment to the nation and its citizens. They emphasize the importance of
# self-government and active citizenship, encouraging everyone to participate in the democratic process. In contrast,
# the other speaker provides a factual update on the weather in Barcelona, which is less inspiring as it
# lacks the emotional and motivational content of the farewell address.
# **Differences:**
# - The farewell address speaker focuses on the values and responsibilities of citizenship, encouraging active participation in democracy.
# - The weather update speaker provides factual information about the temperature in Barcelona, without any emotional or motivational content.
messages = [
user_msg,
AssistantMessage(content=content).to_openai(),
UserMessage(content="Ok, now please summarize the content of the first audio.").to_openai()
]
print(30 * "=" + "USER 2" + 30 * "=")
print(messages[-1]["content"])
print("\n\n")
response = client.chat.completions.create(
model=model,
messages=messages,
temperature=0.2,
top_p=0.95,
)
content = response.choices[0].message.content
print(30 * "=" + "BOT 2" + 30 * "=")
print(content)
from mistral_common.protocol.transcription.request import TranscriptionRequest
from mistral_common.protocol.instruct.messages import RawAudio
from mistral_common.audio import Audio
from huggingface_hub import hf_hub_download
from openai import OpenAI
# Modify OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://<your-server-host>:8000/v1"
client = OpenAI(
api_key=openai_api_key,
base_url=openai_api_base,
)
models = client.models.list()
model = models.data[0].id
obama_file = hf_hub_download("patrickvonplaten/audio_samples", "obama.mp3", repo_type="dataset")
audio = Audio.from_file(obama_file, strict=False)
audio = RawAudio.from_audio(audio)
req = TranscriptionRequest(model=model, audio=audio, language="en", temperature=0.0).to_openai(exclude=("top_p", "seed"))
response = client.audio.transcriptions.create(**req)
print(response)
podman run --rm -it --device nvidia.com/gpu=all -p 8000:8000 \
--ipc=host \
--env "HUGGING_FACE_HUB_TOKEN=$HF_TOKEN" \
--env "HF_HUB_OFFLINE=0" -v ~/.cache/vllm:/home/vllm/.cache \
--name=vllm \
registry.access.redhat.com/rhaiis/rh-vllm-cuda \
vllm serve \
--tensor-parallel-size 8 \
--max-model-len 32768 \
--enforce-eager --model RedHatAI/Voxtral-Mini-3B-2507-FP8-dynamic
# Setting up vllm server with ServingRuntime
# Save as: vllm-servingruntime.yaml
apiVersion: serving.kserve.io/v1alpha1
kind: ServingRuntime
metadata:
name: vllm-cuda-runtime # OPTIONAL CHANGE: set a unique name
annotations:
openshift.io/display-name: vLLM NVIDIA GPU ServingRuntime for KServe
opendatahub.io/recommended-accelerators: '["nvidia.com/gpu"]'
labels:
opendatahub.io/dashboard: 'true'
spec:
annotations:
prometheus.io/port: '8080'
prometheus.io/path: '/metrics'
multiModel: false
supportedModelFormats:
- autoSelect: true
name: vLLM
containers:
- name: kserve-container
image: quay.io/modh/vllm:rhoai-2.25-cuda # CHANGE if needed. If AMD: quay.io/modh/vllm:rhoai-2.25-rocm
command:
- python
- -m
- vllm.entrypoints.openai.api_server
args:
- "--port=8080"
- "--model=/mnt/models"
- "--served-model-name={{.Name}}"
env:
- name: HF_HOME
value: /tmp/hf_home
ports:
- containerPort: 8080
protocol: TCP
# Attach model to vllm server. This is an NVIDIA template
# Save as: inferenceservice.yaml
apiVersion: serving.kserve.io/v1beta1
kind: InferenceService
metadata:
annotations:
openshift.io/display-name: Voxtral-Mini-3B-2507-FP8-dynamic # OPTIONAL CHANGE
serving.kserve.io/deploymentMode: RawDeployment
name: Voxtral-Mini-3B-2507-FP8-dynamic # specify model name. This value will be used to invoke the model in the payload
labels:
opendatahub.io/dashboard: 'true'
spec:
predictor:
maxReplicas: 1
minReplicas: 1
model:
modelFormat:
name: vLLM
name: ''
resources:
limits:
cpu: '2' # this is model specific
memory: 8Gi # this is model specific
nvidia.com/gpu: '1' # this is accelerator specific
requests: # same comment for this block
cpu: '1'
memory: 4Gi
nvidia.com/gpu: '1'
runtime: vllm-cuda-runtime # must match the ServingRuntime name above
storageUri: oci://registry.stage.redhat.io/rhelai1/modelcar-voxtral-mini-3b-2507-fp8-dynamic:1.5
tolerations:
- effect: NoSchedule
key: nvidia.com/gpu
operator: Exists
# make sure first to be in the project where you want to deploy the model
# oc project <project-name>
# apply both resources to run model
# Apply the ServingRuntime
oc apply -f vllm-servingruntime.yaml
# Replace <inference-service-name> and <cluster-ingress-domain> below:
# - Run `oc get inferenceservice` to find your URL if unsure.
# Call the server using curl:
curl https://<inference-service-name>-predictor-default.<domain>/v1/chat/completions
-H "Content-Type: application/json" \
-d '{
"model": "Voxtral-Mini-3B-2507-FP8-dynamic",
"stream": true,
"stream_options": {
"include_usage": true
},
"max_tokens": 1,
"messages": [
{
"role": "user",
"content": "How can a bee fly when its wings are so small?"
}
]
}'
See Red Hat Openshift AI documentation for more details.
Creation
This model was quantized using the llm-compressor library as shown below.
import torch
from transformers import VoxtralForConditionalGeneration, AutoProcessor
from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import QuantizationModifier
# Select model and load it.
MODEL_ID = "mistralai/Voxtral-Mini-3B-2507"
model = VoxtralForConditionalGeneration.from_pretrained(MODEL_ID, torch_dtype=torch.bfloat16)
processor = AutoProcessor.from_pretrained(MODEL_ID)
# Recipe
recipe = QuantizationModifier(
targets="Linear",
scheme="FP8_DYNAMIC",
ignore=["language_model.lm_head", "re:audio_tower.*" ,"re:multi_modal_projector.*"],
)
# Apply algorithms.
oneshot(
model=model,
recipe=recipe,
processor=processor,
)
SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-dynamic"
model.save_pretrained(SAVE_DIR, save_compressed=True)
processor.save_pretrained(SAVE_DIR)
After quantization, the model can be converted back into the mistralai format using the convert_voxtral_hf_to_mistral.py script included with the model.
Evaluation
The model was evaluated on the Fleurs transcription task. Recovery is computed with respect to the complement of the word error rate (WER).