AI Daily · Oct 21, 2026 · Wednesday
BAD INFLUENCE
DeepSeek open-sources DeepInfer, its in-house inference framework
DeepSeek has open-sourced DeepInfer, its in-house inference framework, which it says powers DeepSeek-V4.1-Flash at scale with high throughput and low latency. SGLang congratulates DeepSeek and says it is excited about the large-batch and multi-GPU inference work and plans to bring some of those ideas into SGLang.
Model releases & updates
ElevenLabs BlogOfficial2 more sourcesElevenLabs launches Eleven Realtime, two speech-to-speech models for voice agents
ElevenLabs launches Eleven Realtime v1 Mini and v1 Max, speech-to-speech models that each act as a complete voice agent in more than 90 languages. Artificial Analysis reports 98.8% accuracy for v1 Max and 75.9% for v1 Mini on Big Bench Audio, with v1 Max completing 87.1% of tool-use tasks versus 60.3% for v1 Mini. Developers can use it in ElevenAgents or through a WebSocket API compatible with OpenAI's Realtime API.
Product launches & updates
PyTorch BlogOfficial6 more sourcesGoogle and Meta introduce TorchTPU, a native PyTorch backend for Google TPUs
Google and Meta announce TorchTPU, an open source project that runs PyTorch natively on Google TPUs. The backend supports eager execution and torch.compile, and the team reports validating out-of-box execution on about 5,000 Hugging Face models. Porting a model requires changing model.to("cuda") to model.to("tpu"), though peak performance needs further tuning.
- Google explains how to integrate TorchTPU into TorchTitan and optimize models for TPUGoogle Developers Blog
Mistral releases Voxtral 3 speech models and a voice agent SDK
Mistral AI releases Voxtral 3, a family of speech-to-text and text-to-speech models plus a voice agent SDK that runs listening, reasoning, tool calling, and speaking in one real-time pipeline. The speech-to-text models support over 50 languages, and Voxtral Speech 3 is priced at $13 per 1M characters. The stack is available through Mistral's EU-hosted API or for self-deployment on customers' own infrastructure.
Google Developers BlogOfficial1 more sourceGoogle launches native PyTorch vLLM and SGLang inference engines on TPU
Google announces native PyTorch backends for vLLM and SGLang on TPU, built on TorchTPU and developed with Inferact and RadixArk. The engines add production features such as prefix caching, KV cache offloading, and disaggregated serving, with KV cache transfers handled by the open-source TPU Sync library.
Fireworks AI BlogOfficialFireworks AI makes Meta's Muse Spark Open available on US serverless
Fireworks AI makes Meta's open-weights Muse Spark Open available on its US-only serverless API with no US premium and Zero Data Retention by default. The post reports that the 816B-parameter mixture-of-experts model ranks first of 75 models on Vals AI's Harvey Legal Agent Benchmark at 25.4%, nearly 4x Claude Opus 5's 6.7%. It also reports 53.3% on Rogo's Big Finance Bench, above Claude Opus 5's 49.2%, and 82.9% on Terminal-Bench 2.1 in Meta's evaluations.
Industry news
TechCrunch · AINews1 more sourceAnthropic struggles to secure enough cloud capacity as hyperscalers turn competitors
A leaked S-1 filing reported by The Wall Street Journal says Anthropic has trouble securing enough cloud capacity from AWS, Google, and Microsoft, which increasingly view it as a competitor. Anthropic estimates it could spend more than $750 billion on cloud contracts, including deals still in negotiation, and reported an $80.7 billion net loss in the first half of 2026 on $16.3 billion in revenue. The company is reportedly targeting a November IPO, and it did not respond to TechCrunch's request for comment.
OpenAI@OpenAIOfficial6 more sourcesOpenAI acquires Orchestra Bio to coordinate biopharma drug development
OpenAI says it has acquired Orchestra Bio, a platform that helps biopharma teams coordinate the complex process of drug development. The company plans to build on Rosalind, its AI models and tools for life sciences research, to streamline those operations and help new treatments reach patients sooner.
Papers & research
Anthropic ResearchOfficial2 more sourcesClaude designs protein binders against scorpion neurotoxins, with 17 of 211 lab-confirmed
Anthropic says its internal research model, given a prompt and four follow-ups, designed protein binders against scorpion neurotoxins over 64 hours. Adaptyv Bio and Twist Bioscience tested 211 designs, and 17 met the binding criteria, all against a single toxin, none against several. Anthropic has not tested neutralization in cells or animals, so it does not claim an antivenom, and it is releasing all 211 designs and their data.
Ai2 releases FairStress, a benchmark for hidden demographic asymmetry in LLM decisions
Ai2 introduces FairStress, a benchmark of 300 two-candidate decision scenarios across 29 domains that tests how stable LLM decisions are under adversarial prefixes. The team reports that models switched from stronger candidates to minority-coded ones more often than the reverse, and that counterfactual chain-of-thought reduced this gap while largely preserving accuracy. The benchmark is available for download and is being added to Olmo-eval.
METR BlogOfficial1 more sourceMETR finds Anthropic's misalignment risk report rests on weak covert capability evidence
METR agrees that misalignment risk from Anthropic's covered models is low, but says the report's main reason is limited capability rather than safeguards. METR argues the covert capability benchmarks poorly test whether models could evade Anthropic's monitoring, and that recent incidents weaken several of the report's arguments.
Tips & takes
meng shao@shao__mengX2 more sourcesHow to tune decision models for LLM evals: thresholds, calibration, and cost
Meng Shao summarizes Hamel Husain's post on using decision models as cheap LLM judges for evals. Tuning the threshold instead of keeping the 0.50 default, and checking probability calibration, changed results on 98 real agent traces. The post also weighs F1 against cost, with OpenAI Decisions scoring highest but costing about 4.5 times more than Perplexity.
