Voxtral Small Transcribe 3

Voxtral Small Transcribe 3 is Mistral's flagship speech-to-text, diarization, and audio intelligence model. Released on October 20, 2026, as part of the Voxtral 3 ecosystem, it is designed to bridge the gap between simple acoustic transcription and agentic reasoning over voice. Rather than functioning purely as an automated speech recognition (ASR) pipe, the model processes spoken inputs directly into structured actions, summaries, translations, and function calls across more than 50 languages.

Built using a dense architecture with an undisclosed parameter count, Voxtral Small Transcribe 3 targets high-throughput enterprise speech pipelines and on-premise sovereign deployments. It directly challenges established transcription engines like Whisper large-v3 and proprietary cloud services such as Gemini 3.5 Transcribe. For organizations with strict compliance, legal, or data residency mandates that disqualify multi-tenant cloud APIs, Mistral allows engineers to pull the model onto local compute infrastructure without leaking voice telemetry to third-party endpoints.

What separates this release from legacy transcription architectures is its end-to-end multi-task execution. By combining acoustic processing with a 32,000-token context window, the model eliminates the brittle pipeline of piping raw ASR output into an external language model for post-processing. It handles diarization, keyphrase context biasing, word-level timestamps, and conversational turn detection natively within a single inference pass.

Voxtral Small Transcribe 3 utilizes a dense transformer architecture. Unlike Mixture-of-Experts (MoE) designs that activate a subset of weights per token to minimize compute cycles, a dense model routes all inputs through the entire network parameter set. For speech tasks, dense models provide predictable latency curves and stable memory utilization across varying audio complexities, which is critical when processing real-time streaming audio feeds.

The standout technical metric is its 32,000-token context length. Standard open-source speech models often process audio in rigid 30-second windows with minimal prompt awareness. Voxtral Small Transcribe 3 leverages its expanded context to ingest both temporal audio signals and rich prompt context simultaneously. This enables:

• Keyphrase and lexicon biasing: Developers can feed industry-specific vocabulary, technical acronyms, or speaker directories into the context window before transcription starts.
• Persistent diarization: The model maintains speaker identification consistency across extended multi-speaker conversations rather than losing track of speaker IDs across chunk boundaries.
• Multi-turn audio reasoning: The model holds conversational history in memory, allowing users to issue voice-driven follow-ups, summarization requests, or structured database extractions directly against spoken sessions.

While Mistral maintains an undisclosed parameter footprint for the model weights, its designation as a "Small" tier model in the 2026 lineup positions it within the lightweight enterprise category. It delivers a performance-to-footprint ratio that makes self-hosting viable on workstation-grade silicon.

Voxtral Small Transcribe 3 is built for production audio workflows where transcription accuracy must feed directly into programmatic actions. Key capabilities include:

• Acoustic transcription and translation: Native speech processing across 50+ languages, handling heavy regional accents, background noise, and overlapping speech without requiring auxiliary language identification models.
• Automated speaker diarization: Zero-shot speaker labeling that separates multi-party dialogues (e.g., earnings calls, medical consultations, customer support interactions) with exact start and end boundaries.
• Precise timestamping: Word-level and segment-level timestamps calibrated for subtitle generation, legal discovery synchronization, and compliance archiving.
• End-of-turn detection: Real-time conversational modeling that identifies when a speaker has concluded an utterance, eliminating unnatural pauses in voice agent response loops.
• Direct speech function calling: The model can parse spoken audio and immediately emit valid JSON tool calls. For example, a customer dictating an address change triggers a backend API call without intermediate text extraction scripts.

Typical production use cases include automated call-center auditing, live courtroom and patent office transcription, sovereign clinical documentation, and low-latency voice bots operating inside private corporate networks.

Deploying this model locally requires engineering around Mistral's dense architecture and audio feature extraction pipeline. For developers evaluating Voxtral Small Transcribe 3 hardware requirements, memory bandwidth and dedicated tensor performance will be the primary bottlenecks.

Because Mistral operates with an undisclosed parameter scale, practical local deployments mirror typical 12B to 24B class dense inference setups. Audio processing adds additional activation overhead during feature map extraction. Plan your local AI model undisclosed parameters 2026 hardware budget around the following memory targets:

• 16-bit Float (FP16/BF16): Requires ~32 GB to 40 GB of VRAM. This configuration requires high-end professional hardware like an NVIDIA RTX 6000 Ada, an A100/H100 (40GB/80GB), or an Apple Silicon unified memory workstation.
• 8-bit Quantization (Q8_0 / INT8): Requires ~20 GB to 24 GB of VRAM. This is the baseline for zero-degradation deployment on top-tier consumer hardware.
• 4-bit Quantization (Q4_K_M / INT4): Requires ~12 GB to 16 GB of VRAM. Recommended for standard workstation setups and high-throughput parallel batching.

To run Voxtral Small Transcribe 3 locally with sustainable real-time factors (RTF), the best GPU for Voxtral Small Transcribe 3 setups depends on your deployment profile:

• Consumer Desktop (NVIDIA RTX 4090 24GB): The premier choice for single-card local deployment. Running an 8-bit or 4-bit quantized build gives you full access to the 32k context window without offloading layers to system RAM.
• Apple Silicon (Mac Studio or MacBook Pro with M4 Max / M3 Max, 36GB+ Unified Memory): An excellent local environment for developers building voice agents. Apple's unified memory architecture easily handles dense memory requirements and the 32k token window via Metal-optimized runtimes.
• Multi-GPU Workstations (Dual RTX 4070 Ti Super 16GB or Dual RTX 3090): Ideal for serving both the acoustic encoder and the generation decoders across dedicated PCIe lanes using tensor parallelism.

The best quantization for Voxtral Small Transcribe 3 in mixed agentic workloads is Q4_K_M or AWQ 4-bit. 4-bit quantization retains acoustic phonetic precision while drastically slashing the memory overhead required to sustain concurrent audio streams.

In testing scenarios, Voxtral Small Transcribe 3 performance yields impressive real-time factors. When decoding generated text responses from speech inputs, expected Voxtral Small Transcribe 3 tokens per second range from 45 to 70 tokens/sec on an RTX 4090 under INT4, easily fast enough to maintain natural conversational cadence in voice agent applications.

For quick evaluation, community-supported Ollama builds and vLLM servers provide the fastest path to execution. Once the weights or quantized GGUF/safetensors files are placed in your model directory, configure your Modelfile or vLLM launch parameters:

1# Launching via modern local inference engines
2vllm serve mistralai/voxtral-small-transcribe-3 \
3    --tensor-parallel-size 1 \
4    --max-model-len 32768 \
5    --gpu-memory-utilization 0.90 \
6    --enforce-eager

If utilizing Ollama:

1ollama run voxtral-small-transcribe-3

Ensure your client pipeline feeds raw PCM or standardized 16kHz WAV streams formatted according to the model's audio pre-processor requirements.

Evaluating Voxtral Small Transcribe 3 requires looking at its primary open-source and API counterparts.

OpenAI's Whisper large-v3 has been the standard open-weights transcription baseline since its release. However, architectural differences between the two are significant:

• Context and Reasoning: Whisper is an encoder-decoder network restricted to a 448-token context window that processes audio in 30-second segments. Voxtral Small Transcribe 3 provides a 32,000-token context window. While Whisper merely outputs text, Voxtral can summarize, answer questions, or call functions directly on the audio track.
• Biasing: Injecting custom domain vocabulary into Whisper is notoriously difficult and prone to hallucination loops. Voxtral handles keyphrase prompting natively through its large input prompt space.
• Compute Footprint: Whisper large-v3 runs on roughly 1.55 billion parameters, making it substantially smaller and easier to run on entry-level GPUs (8GB VRAM). Voxtral Small Transcribe 3 has significantly higher hardware demands, requiring 16GB to 24GB VRAM for comfortable quantized inference.

Select Whisper large-v3 if you only need straightforward, low-cost subtitle generation on edge devices with minimal VRAM. Choose Voxtral Small Transcribe 3 if your software architecture demands speaker diarization, direct JSON tool calling from voice, zero-shot multilingual translation, and on-premise execution where the model must think about the audio it just heard.