Three open releases in one week: inference, voice and chips
On 20 October DeepSeek open-sourced DeepInfer, the inference framework behind its production service for DeepSeek-V4.1-Flash. It is Apache 2.0 and runs on Nvidia SM100 GPUs and Huawei Ascend 950 chips. The release covers offline inference; the KV cache system and load balancing needed for a full service are not included.
The same week, Mistral released Voxtral 3: speech-to-text in more than 50 languages, text-to-speech, and a voice agent SDK built on LiveKit. It is available through Mistral's API or for self-deployment.
Also on 20 October, Google open-sourced TorchTPU, a native PyTorch backend for TPUs co-developed with the PyTorch team at Meta. AWS announced TorchNeuron, a native PyTorch backend for Trainium, the same day.
Together these could make it cheaper to run agents yourself: cheaper serving, voice that works on your own hardware, and less lock-in to one chip vendor.
- deepseek-ai/DeepInfer, GitHub: https://github.com/deepseek-ai/DeepInfer
- Mistral, Voxtral 3 announcement: https://mistral.ai/news/voxtral-3/
- The Decoder, 20 Oct 2026: https://the-decoder.com/mistral-ai-launches-voxtral-3-a-full-voice-agent-stack-that-listens-thinks-and-talks-in-real-time/
- Google Cloud Blog, 20 Oct 2026: https://cloud.google.com/blog/topics/ai-infrastructure/torchtpu-is-now-open-source-run-native-pytorch-on-tpus
- PyTorch Blog, TorchNeuron, 20 Oct 2026: https://pytorch.org/blog/introducing-torchneuron-native-pytorch-backend-for-aws-trainium/