World Wire

Tencent’s Hy team released Hy3 . Hy3 is a 295B-parameter Mixture-of-Experts (MoE) model. It activates only 21B parameters per token. The weights ship under the Apache License 2.0 . Hy3 is aimed at reasoning, agentic workflows, and long-context tasks. What is Hy3? Hy3’s architecture contains a sparse MoE with 192 experts and top-8 routing. Only 8 experts fire per token, so compute stays low. The model also uses a Multi-Token Prediction (MTP) layer. MTP predicts several tokens at once for faster decoding. Both vLLM and SGLang enable it through speculative decoding. Property Value Architecture Mixture-of-Experts (MoE) Total parameters 295B Activated parameters 21B MTP layer parameters 3.8B Layers (excluding MTP) 80 MTP layers 1 Attention heads 64 (GQA, 8 KV heads, head dim 128) Hidden size 4096 Intermediate size 13312 Context length 256K Vocabulary size 120832 Experts 192 experts, top-8 activated Supported precisions BF16 A separate Hy3-FP8 checkpoint is ...

OpenAI has released two new Realtime models in its API. They are named gpt-realtime-2.1 and gpt-realtime-2.1-mini . Both target low-latency voice and multimodal experiences. The mini model is the notable part of this release. It is a mini reasoning model for realtime voice. It ships at the same cost as the earlier gpt-realtime-mini . OpenAI also reduced p95 latency by at least 25% across Realtime voice models. That reduction comes from improved caching. What is GPT-Realtime-2.1-mini gpt-realtime-2.1-mini is a mini reasoning model for realtime voice interactions. It responds to audio and text inputs over a live connection. OpenAI positions it as the faster, more cost-efficient option in the lineup. The Realtime API processes and generates audio through a single model. This avoids chaining separate speech-to-text and text-to-speech systems. That single-model design reduces latency and preserves nuance in speech. Reasoning is the main capability here. It means the model...

In this tutorial, we build an end-to-end autonomous AI co-scientist workflow for next-generation EGFR inhibitor discovery, focusing on the C797S osimertinib-resistance mutation in non-small cell lung cancer. We start by resolving the biological target through ChEMBL and UniProt, then mine curated EGFR IC50 bioactivity records and convert them into a clean pIC50 modeling dataset. We use RDKit to standardize molecules, remove salts, aggregate replicate measurements, compute Morgan fingerprints, extract physicochemical descriptors, and analyze scaffold diversity so that our model learns from chemically meaningful representations rather than raw strings. From there, we train a scaffold-split Random Forest QSAR model, evaluate its ability to generalize to unseen chemotypes, interpret potency-driving features with SHAP or model importances, and visualize influential molecular substructures. Finally, we move beyond prediction into generative design by recombining BRICS fragments from potent ...

Synthetic Sciences has released OpenScience , an open-source AI workbench for scientific research. It is licensed under Apache 2.0 and runs on your own infrastructure. The research team frames it as an open alternative to Anthropic’s Claude Science , launched in late June 2026. The pitch is direct. Scientific AI tooling should not be owned by one vendor. OpenScience keeps the workflow open, the models swappable, and the data local. It is an independent project, not affiliated with or endorsed by Anthropic. TL;DR OpenScience is an Apache-2.0, model-agnostic AI workbench for machine learning, biology, physics, and chemistry. It runs the full loop: literature, hypothesis, code, experiment, analysis, and write-up. Any model works (Claude, GPT, Gemini, GLM, Kimi, DeepSeek, local fine-tunes); switching is per-request. It ships 250+ editable skills, plus databases (UniProt, PDB, ChEMBL, arXiv, and ~30 more) as agent tools. It runs on your infrastructure wit...