The Day in AI

• Google released TorchTPU, an open-source PyTorch backend that makes the TPU a native device, plus vLLM and SGLang engines built on it.
• Mistral AI launched Voxtral 3, a speech-to-text model, a text-to-speech model and a voice agent SDK, with transcription from $0.0035 per minute.
• DeepSeek open-sourced DeepInfer under Apache 2.0, the inference framework behind its production services, with launch support for DeepSeek-V4.1-Flash.
• Inception Labs released Mercury 3 on 19 October, a diffusion language model it says runs at 1,068 tokens per second.
• US government ordered federal agencies to remove DeepSeek from their systems, according to TechCrunch.

Until now, PyTorch users who wanted TPUs had to rewrite their code in JAX or go through the older torch_xla library. TorchTPU adds a "tpu" device to PyTorch itself. Code runs eagerly or compiles through torch.compile, which hands the graph to Google's XLA compiler, and multi-chip communication goes through the standard torch.distributed library. Google also shipped vLLM and SGLang serving engines built on TorchTPU, and a TPU fork of TorchTitan for training.

Google reports that its two engines have the lowest cost on the InferenceX benchmark for Qwen 3.5 on Ironwood TPUs, at 2.01 cents per million tokens. Google also reports more than 700 tokens per second to a single user for Kimi K3, a 2.8 trillion parameter model, using a custom Pallas kernel in vLLM-TorchTPU co-developed with Inferact. Both figures come from Google's blog and have no independent check yet.

Mistral AI launched Voxtral 3, a full voice agent stack. Voxtral Small Transcribe 3 transcribes more than 50 languages and handles speaker separation, custom vocabulary and detection of when a speaker has finished. Voxtral Mini TTS 3 speaks over 35 languages and 50 dialects with voices that adjust their emotional tone. The SDK, built on LiveKit, runs listening, reasoning, tool calls and speech in one streaming pipeline, so an agent can act on outside systems during a live call. It's a closed API, and Mistral prices transcription from $0.0035 per minute.

DeepSeek open-sourced DeepInfer, the serving framework behind its own API. It bundles DeepSeek's FlashMLA, DeepGEMM and DeepEP kernels with expert parallelism, draft-token decoding and deterministic sampling, and it started as a fork of vLLM before heavy rework. Launch support covers DeepSeek-V4.1-Flash, the model it updated on 17 October, on NVIDIA SM100 chips and Huawei Ascend 950. This release covers offline inference only, so the KV cache system, the split of prefill and decode across machines, and load balancing across replicas are still private.

• Inception Labs released Mercury 3 on 19 October, a diffusion model that refines a block of tokens in parallel, and says it matches GPT-6 Luna and Claude Haiku 5.5 at 4 to 8 times their speed, priced at $0.20 per million input tokens and $0.70 per million output tokens.
• Ant Group open-sourced five Ming-3.0 speech models, including a mixture-of-experts text-to-speech model with 7.9 billion total and 1.3 billion active parameters under the MIT license.
• OpenAI added a 12-hour prompt caching option to the Responses API, with cache writes costing 2.5 times the input token price.
• ACE Robotics, founded by SenseTime co-founder Xiaogang Wang, released ACE-Brain-0.6 on 19 October and reports it beats GPT-6 Astra on six of eight robot spatial grounding benchmarks.
• US government ordered federal agencies to remove DeepSeek from their systems, according to TechCrunch, and Forkast reports the exclusion extends to products built on the model.
• OpenAI acquired Orchestra Bio, which builds agents for managing drug development, and Axios reports the technology will go into its life sciences suite Rosalind.
• Daytona, which builds cloud infrastructure for AI agents, raised a $60 million Series B led by FirstMark at a $1 billion valuation, according to Fortune.
• Nathan Chen Guangyu, a 17-year-old researcher at Moonshot AI, is joining Discovery Loop, a start-up reportedly co-founded by Jeff Dean, according to an anonymous talent-tracking account, and the company hasn't confirmed it.