← All updates

Mistral Ships Arabic Realtime Voxtral

Mistral's new 4.4B streaming speech-to-text model targets Arabic dialects and mid-sentence language switching

Hi!

DIT Model Radar tracks new AI models daily: real models or confirmed codenames only, no tools, no hype.

Mistral published Voxtral Mini 4B Realtime Arabic on its official Hugging Face organization, a 4.4B-parameter streaming speech-to-text model for Arabic dialects and Modern Standard Arabic. At a configurable 480 ms transcription delay it averages 8.82% character error rate across seven Arabic benchmarks, running through vLLM or Transformers.

What changed: Weights for Voxtral Mini 4B Realtime Arabic are now live in Mistral's Hugging Face repo, fine-tuned from Voxtral-Mini-4B-Realtime-2602 with BF16 weights. The model card documents both vLLM realtime serving and native Transformers support.

Status: Published on Mistral's own Hugging Face organization, with weights and model card live

  • Official Mistral repo hosts the weights, roughly 4.4B parameters in BF16
  • 480 ms configurable delay, 8.82% average CER across seven Arabic benchmarks
  • Usage documented for vLLM realtime serving and Transformers 5.2.0 or newer
  • Not published on the card: any hosted API or pricing for this model

It is a published weights release for Arabic streaming speech recognition, co-developed with Morocco's MTNRA as a sovereign AI asset.

Read more →

[ $1 LAUNCH SPECIAL ]
Voxtral Mini Arabic Realtime on DIT
15× the tokens. Just $1.
≈150M input or ≈38M output tokens.*
Unlocks at launch. *Illustrative from voxtral-small-2507: $0.1/M in + $0.4/M out.

Top AI models. Lower costs.

Access GPT, Claude, Gemini, and more through one OpenAI-compatible API—and save up to 70% on AI API costs. Create your DIT account, get your API key, and start building.