The AI log, open weights · agents and coding · frontier
14 Nov 2026
Perplexity Decider, Claude Meetings, Kitten TTS 2, SOV-Transcribe-1
Open decision models from Perplexity, Claude joining video calls, an 82M speech model for CPUs and Salesforce's open-weight transcription.
By Ali Asif, published
What shipped
- Perplexity
perplexity-ai/pplx-decider-v1.3
Apache-licensed decision models, a 27B-based Decider and a distilled 4B-based Decider Flash, that return probabilities for several questions about one input in a single forward pass, plus a new locate feature for extraction.
- Anthropic
Introducing Claude Meetings, now in beta
Claude joins a Google Meet call as a visible participant, answers questions, builds docs and whiteboards live and writes notes and action items, in beta on Team and Enterprise plans.
- Kitten ML
KittenML/kitten-tts-2-mini
An 82M-parameter Apache-licensed text-to-speech model with 16 languages and 118 voices, shipped as an 82.7 MB int8 ONNX file that runs on CPU.
- Salesforce
salesforce-sovereign-intelligence/SOV-Transcribe-1
An Apache-licensed speech recognition model with 26.6B total and 4.6B active parameters that the card says ranks first among open-weight non-streaming models on Artificial Analysis.
Four primary pages I opened and read today: an open classifier family, a meeting assistant from a frontier lab and two speech models. The numbers below come from those pages, not from my own testing.
Perplexity Decider v1.3
Perplexity released Decider v1.3 and Decider Flash v1.3 under the Apache license, on Hugging Face, OpenRouter and its Decisions API. A decision model is a general-purpose classifier. You give it an input and a set of questions, each with possible answers, and it returns a probability distribution per question. Question types are yes or no, multiple choice and ordered scores.
The design is the interesting part. Perplexity adapts a generative LLM backbone and keeps only the output rows for the label tokens, so one forward pass answers several questions about the same input with no token-by-token generation. The larger model is based on Qwen3.8-27B, and the smaller Flash model is distilled from it into Qwen3.5-4B. Training runs in three stages: generalist full fine-tuning, a targeted LoRA stage and a vision step.
Version 1.3 adds bidirectional attention with isolated questions, so each question sees the shared state and its own tokens only. Perplexity says this gives more stable answers than letting every question see every other one. It also adds locate, which returns the exact matching text from the input with character offsets, aimed at entity recognition and extractive question answering.
Claude Meetings
Anthropic put Claude Meetings in beta for Team and Enterprise plans, starting with Google Meet. Claude joins as a visible participant and posts a notice in the call chat. People can ask it questions out loud or in chat, and it answers in writing by default. It can build a doc or whiteboard live while teammates edit, then writes notes and action items.
The privacy section is what I read most closely. Claude stays out of calls with outside guests by default, uses each person's connected tools only with that person's permission, and saves a transcript, summary and screen descriptions rather than audio. Some organizations are excluded during the beta, including those with zero data retention or HIPAA-ready setups.
Kitten TTS 2 Mini
Kitten ML's model has 82M parameters, 16 languages and 118 voices at 24 kHz. The int8 ONNX file is 82.7 MB and runs on CPU without PyTorch. It is Apache licensed, and the card shows a short Python usage example. A small speech model that needs no GPU is useful for local tools, though I have not listened to the voices.
SOV-Transcribe-1
Salesforce's open-weight speech recognition model pairs a Whisper large-v3 audio encoder with Gemma 4 26B A4B, 26.6B parameters in total and 4.6B active. The card says it ranks first among open-weight non-streaming models on Artificial Analysis and second of 23 systems on VoiceCodeBench as of November 12. It supports seven languages, style control through plain-language instructions and context biasing for names and technical terms.
What I'd watch next
Whether Decider holds up on my own routing and judging tasks, and whether the vLLM and Transformers support for SOV-Transcribe-1 lands upstream as the card says it will.