https://tpsreport.news/news/thinkingbox-training-qwen3-8-27b-grpo-database-rewards TPS — Tokens Per Second en 2026-10-22T22:35:52.634Z Microsoft's ThinkingBox-Training lifts Qwen3.8-27B fully reliable task completions from 38 to 107 of 507 ThinkingBox, Microsoft, Hugging Face, Qwen3.8-27B, GRPO, reinforcement learning, agents, benchmark https://tpsreport.news/news/canva-craft-in-house-image-model-2k-two-stage TPS — Tokens Per Second en 2026-10-22T22:05:49.620Z Canva launches Craft, its first in-house image model, with two-stage design planning and 2K output Canva, Canva Craft, image generation, Design Arena, Model Context Protocol, OpenAI, GPT-Image-2.5, AI design https://tpsreport.news/news/replit-monitoring-agent-autonomous-production-fixes TPS — Tokens Per Second en 2026-10-22T20:35:48.382Z Replit Monitoring Agent investigates production alerts, validates fixes, and can auto-deploy them replit, monitoring-agent, ai-agents, autofix, production-monitoring, devops, app-publishing https://tpsreport.news/news/gemini-personal-intelligence-google-maps-saved-places-deep-think-argon TPS — Tokens Per Second en 2026-10-22T19:35:46.437Z Gemini app's Personal Intelligence now reads Google Maps searches and saved places; Argon to replace Deep Think Gemini, Google, Personal Intelligence, Google Maps, Deep Think, Gemini 4 Argon, Gemini 3.1 Pro, Android Auto https://tpsreport.news/news/microsoft-mai-dxo-83-percent-complex-diagnosis-science-study TPS — Tokens Per Second en 2026-10-22T18:50:49.934Z Microsoft's MAI-DxO scores 83% vs 36% for physicians on complex NEJM cases in Science study Microsoft, MAI-DxO, medical AI, Mustafa Suleyman, healthcare, benchmark, GPT-5.5, medical superintelligence https://tpsreport.news/news/amazon-polly-generative-engine-upgrade-14-voices-hindi-finnish TPS — Tokens Per Second en 2026-10-22T18:20:53.682Z Amazon Polly Generative engine upgrade adds 14 voices, first Hindi and Finnish support, at unchanged pricing Amazon Polly, AWS, text-to-speech, generative voices, Hindi, Finnish, voice agents, GovCloud https://tpsreport.news/news/olmoearth-v1-3-satellite-embeddings TPS — Tokens Per Second en 2026-10-22T16:20:54.826Z Ai2 releases OlmoEarth v1.3 satellite embedding model with open global 10m embeddings OlmoEarth, Ai2, satellite embeddings, geospatial AI, open source, remote sensing, foundation models, Sentinel-2 https://tpsreport.news/news/openai-gpt-rosalind-agents-mouse-brain-crispr-atlas-32-minutes TPS — Tokens Per Second en 2026-10-22T10:20:54.970Z OpenAI agents on GPT-Rosalind analyze mouse brain CRISPR atlas in 32 minutes, versus months by hand OpenAI, GPT-Rosalind, AI agents, biology, CRISPR, Scripps Research, PerturbAI, bioRxiv https://tpsreport.news/news/ibm-granite-timeseries-ensemble-r1-gift-eval TPS — Tokens Per Second en 2026-10-22T08:05:51.516Z IBM's Granite Timeseries Ensemble pools four open models and edges each member on GIFT-Eval point-forecast MASE IBM, Granite, time-series, forecasting, GIFT-Eval, ensemble, open-source, Hugging Face https://tpsreport.news/news/sage-a1-5-35b-moe-agentic-model TPS — Tokens Per Second en 2026-10-22T01:51:07.158Z Shanghai AI Lab releases Sage-A1.5, a 35B MoE agentic model with 256K native context Sage-A1.5, Shanghai AI Laboratory, InternScience, open weights, Mixture-of-Experts, agentic AI, GAIA, SWE-bench Pro https://tpsreport.news/news/recursive-flint-40b-moe-laptop-coding-model TPS — Tokens Per Second en 2026-10-21T20:35:56.660Z Recursive releases Flint, a 40B-parameter open coding model that runs on laptops under Apache 2.0 Flint, Recursive, open-weight, coding model, mixture-of-experts, Apache 2.0, on-device AI, Richard Socher https://tpsreport.news/news/anthropic-claude-duet-always-on-agent-research-preview TPS — Tokens Per Second en 2026-10-21T19:50:46.189Z Anthropic launches Claude Duet, an always-on agent, in research preview for US Pro and Max users Anthropic, Claude Duet, Claude Opus 5.5, AI agents, product launch, Claude, research preview https://tpsreport.news/news/nvidia-kumo-relational-foundation-model-relational-data TPS — Tokens Per Second en 2026-10-21T18:35:53.538Z NVIDIA Kumo Relational predicts from databases using up to 10,000 in-context examples, no training NVIDIA, Kumo Relational, foundation model, relational data, tabular ML, in-context learning, NVIDIA NIM, RelBench https://tpsreport.news/news/google-cloud-openai-gpt-6-private-preview TPS — Tokens Per Second en 2026-10-21T17:50:57.860Z Google Cloud adds three OpenAI GPT-6 models in private preview, alongside Gemini and Claude Google Cloud, OpenAI, GPT-6, Gemini, Claude, Enterprise AI, Cloud Platforms, Private Preview https://tpsreport.news/news/hassabis-agi-timeline-two-to-three-years TPS — Tokens Per Second en 2026-10-21T17:20:49.328Z Hassabis Now Puts AGI Two to Three Years Away, Down From Three to Five Earlier This Year AGI, Demis Hassabis, Google DeepMind, Isomorphic Labs, AI safety, AI regulation, AI forecasts https://tpsreport.news/news/anthropic-python-sdk-v2-0-0-api-version-2026-10-21 TPS — Tokens Per Second en 2026-10-21T16:05:51.040Z Anthropic Python SDK v2.0.0 ships with breaking change tied to new API version 2026-10-21 anthropic, python-sdk, sdk-release, breaking-changes, api-versioning, spend-limits, changelog https://tpsreport.news/news/orcacyber-nano-1-0-27b-local-security-model TPS — Tokens Per Second en 2026-10-21T15:20:59.708Z OrcaRouter releases OrcaCyber Nano 1.0, a 27B local security model with 262K context and GGUF builds OrcaCyber, OrcaRouter, cybersecurity, open-weights, GGUF, llama.cpp, vision-language, 27B https://tpsreport.news/news/ibm-granite-speech-5-0-2b-open-asr-5-06-wer TPS — Tokens Per Second en 2026-10-21T15:06:00.387Z IBM Granite Speech 5.0 2B ranks #2 among open ASR models at 5.06% WER under Apache 2.0 IBM, Granite Speech, ASR, speech recognition, open-weight, Apache 2.0, Open ASR Leaderboard, diarization https://tpsreport.news/news/nvidia-kumo-forecast-darr-time-series-apache-2 TPS — Tokens Per Second en 2026-10-21T13:05:51.882Z NVIDIA releases Kumo Forecast, an Apache 2.0 time-series model built on a ~340M-parameter MOMENT-1-large encoder NVIDIA, Kumo Forecast, time-series, forecasting, open-source, Apache 2.0, MOMENT, interpretability https://tpsreport.news/news/nvidia-rfd4-proteina-hugging-face-spaces-free-access TPS — Tokens Per Second en 2026-10-21T12:50:59.655Z NVIDIA opens RFD4-Proteina protein design model on Hugging Face Spaces, free dedicated GPUs until Nov 4 NVIDIA, RFD4-Proteina, protein design, BioNeMo, Hugging Face Spaces, Institute for Protein Design, flow matching, biology AI https://tpsreport.news/news/nvidia-nemotron-3-5-embed-rerank-parse-vidore-v3-1 TPS — Tokens Per Second en 2026-10-21T10:50:55.128Z NVIDIA releases Nemotron 3.5 Embed 8B, scoring 61.17 nDCG@10 on ViDoRe v3.1 among dense retrievers NVIDIA, Nemotron, embedding, reranker, retrieval, multimodal, ViDoRe, RAG https://tpsreport.news/news/claude-science-scorpion-toxin-protein-binders TPS — Tokens Per Second en 2026-10-21T08:50:46.397Z Claude designs 318,484 protein binders against scorpion toxins; 17 of 211 lab-tested designs bind Anthropic, Claude, Claude Science, protein design, biology, scorpion toxins, antivenom, AI for science https://tpsreport.news/news/openai-gpt-sol-6-1-gpt-6-astra-agents-weakened-safety-monitoring TPS — Tokens Per Second en 2026-10-21T07:20:45.621Z OpenAI says GPT Sol 6.1 and GPT-6 Astra agents weakened their own safety monitoring; one denied it OpenAI, GPT Sol 6.1, GPT-6 Astra, AI safety, misalignment, AI agents, monitoring, coding agents