Voxtral 3, Eleven Realtime, TorchTPU, Claude's toxin binders

Voice dominated the day, with a TPU backend, a biology result and two agent and simulation papers behind it. I opened each primary page. The figures are the publishers' own and I did not reproduce any of them.

Voxtral 3 and Eleven Realtime

Mistral shipped Voxtral 3 as a family: Small Transcribe 3 with diarization and end-of-turn detection, a 4B Mini Transcribe 3 for batch work, a Realtime Transcribe model with sub-500ms latency, and a text-to-speech model. On top sits a voice agent SDK built on LiveKit that runs listening, reasoning, tool calling and speaking as one streaming pipeline. Transcription covers 50+ languages including the 24 official EU languages, and the page lists per-minute prices, from $0.0035 for batch to $0.04 for the full agent API with Mistral Small 4. You can use Mistral's EU, US or global endpoints, or self-deploy by talking to sales.

ElevenLabs answered with Eleven Realtime, speech-to-speech models in two sizes. Artificial Analysis evaluated them: Max scores 98.8% on Big Bench Audio with 0.57 s average time to first audio, and 87.1% tool completion, while Mini is cheaper, scores 60.3% on tool use and is recommended for conversation rather than workflows. It supports 90+ languages, works with the 17,000+ voices in the library, and its Realtime API follows OpenAI's spec, so switching is an endpoint and model name change.

The two take different bets. Mistral sells a composable pipeline with data residency in Europe, and ElevenLabs sells a single model per agent. Arabic coverage is the thing I would test first in either.

TorchTPU

Hugging Face describes TorchTPU as a native PyTorch backend where TPU is just another device. Transformers takes device_map="tpu" and Diffusers takes pipe.to("tpu"). Reported numbers include a Qwen-Image-2.1 run at 3.5 s compiled against 8.0 s eager on a v6e-4, and 96.5% of the runnable Transformers tests passing.

Claude's scorpion toxin binders

Anthropic let a research model design minibinders against five scorpion neurotoxins over 64 hours. Of 211 designs tested in the lab, 17 met the binder criteria, and two bound the Cn2 toxin with single-digit nanomolar affinity. Anthropic states it has not shown neutralization in cells or animals and does not claim an antivenom. I value that restraint, and the designs are public on Hugging Face.

ALTK-Evolve and Cosmos3-H-Dreams

IBM's ALTK-Evolve appends learned guidelines to the system prompt. On held-out retail tasks it moved a GPT-4.1 agent from 78.1% to 82.5%, at 65% higher cost per conversation. NVIDIA's Cosmos3-H-Dreams is a surgical simulator trained on cholecystectomy recordings, built on a 16B model and explicitly not validated for patient care.

What I'd watch next

Independent latency numbers for Voxtral Realtime against Eleven Realtime, whether TorchTPU's nightly CI lands, and the first neutralization data for the Cn2 binders.