Voxtral-Mini-4B-Realtime for real-time speech transcription is now available in Amazon SageMaker JumpStart
Deploy Mistral's streaming speech model inside your own AWS account — real-time, multilingual, with tunable latency vs. accuracy trade-offs.
View original announcement →Visual Summary
graph TD
A{{Voxtral-Mini-4B-Realtime}}:::announced
B(SageMaker JumpStart):::compute
C(SageMaker Studio):::compute
D(SageMaker Python SDK):::compute
E([Real-Time Streaming]):::feature
F([Multilingual 13 Languages]):::feature
G([Configurable Latency]):::feature
H(Amazon CloudWatch):::storage
I((Developer)):::external
I ==>|"deploys"| B
B ==>|"hosts"| A
C -->|"one-click deploy"| B
D -->|"programmatic deploy"| B
A -->|"transcribes"| E
A -->|"supports"| F
A -.->|"tunes"| G
A -->|"logs to"| H
classDef announced fill:#ff9900,stroke:#ec7211,color:#fff,font-weight:bold
classDef compute fill:#e3f2fd,stroke:#1565c0,color:#1565c0
classDef storage fill:#e8f5e9,stroke:#2e7d32,color:#2e7d32
classDef feature fill:#fff3e0,stroke:#e65100,color:#e65100
classDef external fill:#f5f5f5,stroke:#616161,color:#616161
What's New
AWS has made Voxtral-Mini-4B-Realtime-2602, a multilingual real-time speech transcription model from Mistral AI, available in Amazon SageMaker JumpStart. The model features a natively streaming architecture designed for low-latency audio-to-text transcription across 13 languages. Customers can deploy it directly from SageMaker Studio or via the SageMaker Python SDK with minimal setup.
How It Works
- Streaming-native architecture: The model is built from the ground up for real-time transcription, processing audio as a continuous stream rather than in discrete batches, enabling sub-second latency responses.
- 4B parameter scale: At 4 billion parameters, the model balances transcription quality with computational efficiency, making it suitable for deployment on standard SageMaker inference instances.
- Configurable transcription delay: Users can tune the trade-off between latency and accuracy by adjusting transcription delay parameters, allowing the model to emit partial transcripts sooner or wait for more audio context before committing.
- Multilingual support: The model natively handles transcription across 13 languages without requiring separate language-specific models or explicit language-detection preprocessing.
- SageMaker JumpStart deployment: The model is accessible via the SageMaker Studio Models landing page or programmatically through the SageMaker Python SDK, deploying to a managed real-time inference endpoint in the customer's AWS account.
- Foundation model hosting: Once deployed, the endpoint follows standard SageMaker real-time inference patterns, supporting autoscaling, logging via CloudWatch, and integration with other AWS services.
Why It's Important
- Low-latency speech applications become accessible: Developers can now build real-time captioning, live transcription, and voice-driven interfaces on AWS without managing complex streaming inference infrastructure from scratch.
- Mistral AI's frontier audio model on AWS: Voxtral represents Mistral AI's entry into audio AI, and its availability on JumpStart brings a competitive, non-AWS-native model into the managed AWS ecosystem, expanding customer choice.
- Reduces time-to-deployment: JumpStart's one-click deployment eliminates the need to containerize, configure, and optimize the model manually, cutting deployment time from days to minutes.
- Configurable latency-accuracy trade-off: This is practically significant for use cases like live broadcasting (prioritize speed) versus medical transcription (prioritize accuracy), giving teams a single model that adapts to different SLA requirements.
- Multilingual coverage in one endpoint: Supporting 13 languages from a single deployed model reduces infrastructure overhead for global applications compared to maintaining separate per-language models.
How It's Different
- Natively streaming vs. batch-first models: Unlike models such as OpenAI Whisper (which processes fixed audio chunks) or Amazon Transcribe's standard mode, Voxtral-Mini-4B-Realtime is architecturally designed for continuous streaming, not retrofitted for it.
- Configurable delay parameter: Most transcription services offer fixed latency profiles; Voxtral's explicit delay configuration gives engineers fine-grained control over the latency-accuracy curve at inference time.
- Mistral AI provenance: This is a third-party frontier model (not an AWS-built model), distinguishing it from Amazon Transcribe and giving customers access to Mistral's research lineage and model characteristics.
- Self-hosted on customer infrastructure: Unlike managed API services (Amazon Transcribe, Deepgram, AssemblyAI), the model runs inside the customer's AWS account, keeping audio data within their VPC and security boundary.
- 4B parameter efficiency: Compared to larger speech models, the 4B parameter footprint targets a cost-performance sweet spot suitable for real-time workloads without requiring high-end GPU instances.
When to Prefer It
- Live captioning and accessibility: When building real-time closed captioning for video conferencing, live events, or broadcast media where sub-second latency is a hard requirement.
- Data residency and compliance requirements: When audio data cannot leave the customer's AWS account or VPC, making managed third-party transcription APIs unsuitable.
- Multilingual contact center analytics: When processing customer calls across multiple languages in real time and a single unified model is preferable to routing to language-specific services.
- Voice-driven agentic applications: When integrating speech input into LLM-based agents or real-time assistants where streaming transcription must feed downstream NLP pipelines with minimal delay.
- Latency-accuracy tuning is needed: When different deployment environments (e.g., demo vs. production vs. medical) require different latency profiles from the same underlying model.
- Cost-sensitive real-time workloads: When a 4B parameter model provides sufficient accuracy and the team wants to avoid the per-minute pricing of managed transcription APIs at high call volumes.
Availability
- GA status: Generally available as of July 13, 2026, via Amazon SageMaker JumpStart.
- Access method: Available through the SageMaker Studio Models landing page (SageMakerPublicHub) or the SageMaker Python SDK; no separate sign-up required beyond standard AWS account access.
- Regional availability: Specific supported regions are not enumerated in the announcement; customers should check the SageMaker JumpStart model catalog within their target region for availability.
- Pricing model: Standard SageMaker real-time inference pricing applies based on the instance type selected for the endpoint; there is no separate per-minute transcription charge, but Mistral AI license terms must be reviewed and accepted before use.
- License compliance: As with all third-party JumpStart models, customers are responsible for reviewing and complying with Mistral AI's applicable license terms before deploying or using the model.
- Existing endpoint continuity: Consistent with JumpStart policy, any endpoints deployed from this model will remain functional even if the model is later delisted from the catalog.
