Voxtral Mini 3B 2507 (Amazon Bedrock)

Voxtral Mini 3B 2507 is the smaller sibling in Mistral's Voxtral audio-language family, designed to handle spoken-input tasks such as transcription and audio-conditioned question answering while remaining light enough for resource-constrained deployments. An architecture-encyclopedia breakdown describes it as a conditional-generation model built on a VoxtralForConditionalGeneration backbone that pairs a dedicated audio tower with a language stack. The audio side uses a Conv1d-based VoxtralEncoder with positional embeddings, followed by 32 VoxtralEncoderLayer modules, each containing self-attention with 20 heads of head dimension 64 together with MLP and normalization sublayers, while the overall parameter count sits at roughly 4.68 billion in bfloat16 precision.

The combination of a compact ~4.68B parameter footprint, bfloat16 weights, and a layered audio-then-text encoder makes Voxtral Mini 3B 2507 well suited to practical speech workloads where latency, memory cost, and on-device feasibility matter more than raw scale. It is released with open weights, which lets teams fine-tune or self-host the model for domain-specific audio understanding, voice agents, or accessibility tooling rather than relying solely on a hosted API. Within the broader Voxtral line it represents the entry-level option aimed at efficient everyday audio-language tasks, with the larger Voxtral-Small-24B-2507 positioned for higher-capacity use cases that demand greater reasoning depth.