# Mistral Voxtral 3 and Large 4 Multimodal Release

URL: https://3ammarketer.com/news/mistral-voxtral-3-multimodal-model-release
Published: 2026-10-21

> Mistral AI launched Voxtral 3 for real-time speech and Mistral Large 4 with 1T parameters, spanning sub-500ms latency benchmarks and tiered API rates.

Mistral AI released Voxtral 3 on October 20, 2026, combining a voice agent SDK with speech models alongside its 1-trillion parameter multimodal model, [Mistral Large 4](/news/mistral-large-4-release-specs-pricing). The release enables low-latency voice workflows either on Mistral's EU-hosted API or across self-deployed private infrastructure. [1][2]

## What shipped in the October 2026 multimodal release

Voxtral 3 brings listening, thinking, tool calling, and speaking into a single real-time streaming pipeline. The stack supports multi-step agent interactions across more than 50 languages, expanding to low-resource Asian and Indic languages alongside 24 official EU languages. [2]

On the reasoning side, Mistral Large 4 provides the multimodal foundation as an open-weight mixture-of-experts model. It features 1 trillion total parameters with 49 billion active parameters and a 1 million token context window. Teams exploring open architectures can compare these parameters with our [open-weight LLMs tracking](/data/open-models) or assess how token capacity operates in our guide to [context window limits and costs](/blog/what-is-a-context-window-in-ai). [1]

## Transcription, diarization, and sub-500ms latency

The speech stack spans specialized models cataloged under Mistral's Premier tier. Voxtral Small Transcribe 3 performs multilingual transcription and diarization across over 50 languages, using context-biasing to improve accuracy. According to Mistral's release benchmarks, Voxtral Small Transcribe 3 outperforms Whisper, Scribe v2, and Gemini 3.5 Transcribe across diarization and multilingual transcription. [2][3]

For agent systems, Voxtral Realtime Transcribe 3 delivers transcriptions at sub-500ms latency. Voxtral Mini TTS 3 supports 35 or more languages and over 50 dialects. Voxtral TTS 3 adds streaming and voice cloning, while Pro Voice adaptation reproduces an individual's vocal style and emotional range using 30 minutes of audio. [2][3]

## Multimodal API pricing and subscription tiers

Mistral prices voice APIs on duration and characters rather than raw tokens. The 4-billion parameter Voxtral Mini is billed per minute of audio, while Voxtral Speech 3 costs $13 per 1 million characters. In contrast, text and multimodal reasoning with Mistral Large 4 costs $0.70 per million input tokens and $2.10 per million output tokens, as tracked in our [LLM API pricing timeline](/data/model-pricing). [2][5]

Published API rates across speech, vision, and multimodal inference endpoints. [2][5]

| Model or Service | Pricing Metric | Input Rate | Output Rate |
| --- | --- | --- | --- |
| Voxtral Speech 3 | Per 1 million characters | $13 | - |
| Mistral Large 4 | Per 1 million tokens | $1.36 | $4.18 |
| Google Gemini 3.5 Transcribe | Per 1 million tokens | $2 | $12 |
| Ministral 3 3B | Per 1 million tokens | $0.10 | $0.10 |

Output API pricing per million tokens across selected models in October 2026. [5]

- Ministral 3 3B: 0.1 $ per 1M tokens
- Mistral Small 4: 0.6 $ per 1M tokens
- Mistral Large 3: 1.5 $ per 1M tokens
- Mistral Large 4: 4.18 $ per 1M tokens
- Gemini 3.5 Transcribe: 12 $ per 1M tokens

Input API pricing per million tokens across selected models in October 2026. [5]

- Ministral 3 3B: 0.1 $ per 1M tokens
- Mistral Small 4: 0.15 $ per 1M tokens
- Mistral Large 3: 0.5 $ per 1M tokens
- Mistral Large 4: 1.36 $ per 1M tokens
- Gemini 3.5 Transcribe: 2 $ per 1M tokens

Platform access is partitioned across standard subscription plans alongside per-minute voice consumption fees. The Pro plan costs $14.99 per month excluding taxes and includes $25.5 per month in API credits, while the Team tier costs $24.99 per user per month excluding taxes. Enterprise agreements cover custom models, custom agents, audit logs, SAML SSO, and private deployments. [4]

## EU API hosting versus self-deployed infrastructure

Engineering teams can consume Voxtral 3 and Mistral Large 4 via Mistral's hosted endpoints in the European Union or deploy the weights directly on their own infrastructure. Hosted endpoints integrate directly with moderation pipelines like Mistral Moderation 2, which offers a 128k context window and jailbreaking detection. [2][3]

Deploying privately allows organizations to keep voice data entirely within regional boundaries. However, commercial deployments of self-hosted models require a Mistral license with distinct terms for production use and derivatives. [2][4]

Limit: The cited benchmark figures do not include public word-error-rate comparisons across all 50 languages, and Mistral has not published self-hosted hardware throughput benchmarks for the 1-trillion parameter Large 4 architecture.

For conversational voice bots, Voxtral Realtime Transcribe 3 provides sub-500ms streaming latency, while Voxtral Mini offers cheaper batch transcription billed per minute. If strict sovereignty mandates require offline audio ingestion, secure a commercial Mistral license before deploying model weights on private infrastructure. [2][4]

## Sources

1. [New AI Model Releases: LLMs & Open Weights | AI/TLDR](https://ai-tldr.dev/releases/model/)
2. [Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages](https://mistral.ai/news/voxtral-3/)
3. [Models Overview | Mistral Docs](https://docs.mistral.ai/getting-started/models/)
4. [Pricing | Mistral](https://mistral.ai/pricing/)
5. [LLM API pricing timeline (3AM Marketer data set)](https://3ammarketer.com/data/model-pricing)
