The family includes Transcribe models (Mini is 4B parameters), Mini TTS 3 and a LiveKit-based SDK for real-time listening, tool use and speaking. Batch transcription costs from $0.0035 a minute, and the realtime model claims under 500ms latency. Mistral says it leads on multilingual diarization but names no licence for the weights.
Mistral AI
Working for about 64 hours with four human check-ins, an internal research model produced 318,484 protein designs and picked 180 for testing. Two labs confirmed 17 binders, the best of them at single-digit nanomolar affinity. Neutralisation is untested, so Anthropic makes no antivenom claim. All 211 designs are public.
Anthropic
A Peec AI marketer spent about $50 on a site of unverified, model-generated statistics. Google dropped it from search, but ChatGPT kept retrieving it, sending 10,000 sessions, and 330 domains linked to it. If your agent cites numbers it found on the web, check them.
Peec AI
Figma added 20 MCP partners at its first Config in India. Field told TechCrunch the protocol favours the big frontier labs, and that Figma deliberately limits how much design data it exposes to agents because more exposure means more room for misuse.
TechCrunch
The Apache 2.0 project provides custom ATen kernels, torch.compile integration and data- and tensor-parallel training, with examples from Llama to UNet. Moving existing PyTorch code onto TPUs is meant to take a configuration change, not a rewrite. It follows Amazon's TorchNeuron in the race to loosen Nvidia's grip.
GitHubThe 61st release also randomises the direct map's location and adds hibernate on arm64, the RISC-V vector extension, Intel Arc A-series graphics and up to 64 virtual CPUs per vmd guest. The graphics drivers now come from Linux 6.18.
OpenBSD