AI Daily · 2026-10-16

The standout release is Reflection's open 501B-parameter MoE model Beam, available on Hugging Face under Apache 2.0 with a 1M-token context and aimed at coding and agentic workloads. Tencent Hunyuan introduced Hy-Memory-30B-A3B, a memory-writer model that edges out Claude Opus 5 and other larger models on several memory-framework benchmarks. NVIDIA and AI4Bharat open-sourced a roughly 2.8-trillion-token Indic pretraining corpus covering 22 languages, while NVIDIA also expanded its Nemotron Super VL and Bolt edge models. IBM added Granite Switch 4.1, which combines chat, decision-making, and speech-to-text in a single endpoint. Separately, OpenAI extended academic access to GPT-Rosalind, Google DeepMind highlighted Antigravity for antimicrobial-peptide discovery, and NVIDIA launched RTX Spark laptops.

North America · First-hand

OpenAI

⭐⭐ [Product Update] Bringing GPT-Rosalind to academic research labs

OpenAI announced the Rosalind Academic Plan, which offers eligible academic labs sponsored access to GPT-Rosalind. GPT-Rosalind is OpenAI's dedicated life sciences model, combining frontier reasoning with the Rosalind workbench for specialized tool orchestration across medicinal chemistry, genomics, wet-lab assistance, and other scientific applications. Eligible labs can receive up to 30 Business Standard seats at no cost for up to 12 months, with access to GPT-Rosalind subject to trusted-access approval. The post cites feedback from labs including the Zhao Lab at UCSF, the Bourdenx Lab, and Dr. Brian Hie of Stanford University and the Arc Institute.

Google

⭐⭐ [Research] Accelerating the search for antimicrobial peptides with Google Antigravity

A Google DeepMind blog post describes how biologists used Google Antigravity, an agentic development platform, to speed up the search for antimicrobial peptides (AMPs). Antigravity offers an agentic science workbench that compresses a previously fragmented search across multiple databases into a single session, helping biologists discover, structurally evaluate, and prioritize candidate targets quickly. The platform also uses ProtNLM, a deep learning model that predicts biological function from amino acid sequences, to scan millions of previously uncharacterized proteins in the UniProt database. A technical walkthrough shows biologists using Antigravity to uncover a novel antimicrobial peptide candidate predicted to be active against pathogens involved in respiratory tract infections; that lead is now undergoing wet-lab validation.

East Asia · First-hand

Tencent Hunyuan

⭐⭐⭐ [Model Release] tencent/Hy-Memory-30B-A3B

Tencent Hunyuan released Hy-Memory-30B-A3B, a compact long-term-memory writer model with roughly 3B active parameters, licensed under Apache-2.0, designed to execute the write path inside external memory frameworks. It directly replaces the general-purpose LLM inside Mem0, MemOS, and Hy-Memory, covering Chat, Working, and Coding scenarios, deciding what information to retain, how precisely to represent it, and how new evidence updates the persistent memory bank. The model is trained in two stages, Memory On-Policy Distillation followed by Attributed Memory Reinforcement Learning, and transfers directly to unseen memory frameworks such as EverMemOS. On a fixed 20% benchmark subsample across the four harnesses (Mem0, Hy-Memory, MemOS, EverMemOS), it scores a macro average of 55.49, ahead of Claude Opus 5 at 54.62, DeepSeek V4.1 Flash at 53.64, Hy3 at 51.63, and Qwen3.6-35B-A3B at 50.37. Evaluations change only the memory writer, holding the harness version, prompts, retrieval configuration, answer model, and judge fixed.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐⭐⭐ [Model Release] Welcome Beam: a 501B open model you can own, customize, and scale

Reflection, a new US lab, released its first model, Beam, openly on Hugging Face. Beam is a mixture-of-experts model with 501B total parameters and 23B active per token, text-only input, a 1M-token context window, and an Apache 2.0 license, aimed at coding and agentic use cases for teams that want a frontier-scale model on their own hardware, infrastructure, or cloud. It has 52 layers with 256 experts per MoE layer and 8 routed per token, ships in BF16, FP8, and NVFP4 checkpoint formats, and supports five reasoning levels (low, medium as default, high, xhigh, max), reasoning that stays in context across turns via preserve_thinking, and tool calling with a dedicated grammar. Beam has day-0 support in transformers, SGLang, and vLLM, comes with a dedicated Rust-based harness, and has inference, SFT, and RL recipes from Hugging Face and the Reflection team. Reflection is offering $10,000 in Hub credits to the best projects built with Beam.

⭐⭐⭐ [Model Release] Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint

IBM released Granite Switch 4.1, comprising the Granite-Switch-4.1-8B and -30B models, which handle natural-language responses, general and specialized decisions, and speech-to-text tasks from a single hosted endpoint. Each model combines three parts: a new Decision aLoRA for general-purpose decision-making, a specialized Guardian aLoRA that the post says on average outperforms Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions and Clef-Flash on safety-related decision tasks, and an embedded Granite-Speech-5.0-Turbo-CTC, described as the fastest ASR model on the OpenASR Leaderboard. Combined with the base model, all three take up roughly the same amount of GPU VRAM as the base model alone. The post also states that the Switch adapters are cache-efficient Activated LoRAs, so a KV cache created from a base-model generation can be reused for subsequent adapter-based decisions, that users can compose their own Granite Switch models from Granite Libraries, and that an Adapter SDK for training adapters in Mellea is planned.

⭐⭐ [Model Release] Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models

NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter model that the company says delivers the fastest task completion and highest accuracy for specialized always-on agents. Two edge models join it: Nemotron 3.5 Bolt 74B-A9B 24GB, optimized to balance intelligence and speed per GB for local agentic workloads, and Nemotron 3.5 Bolt E4B 4GB. The source text is truncated at this point and provides no further details.

Product Update

⭐⭐ [Product Update] NVIDIA RTX Spark PCs Available Today — A New Beginning for Windows PCs

NVIDIA announced that NVIDIA RTX Spark laptops are available today. The machines combine RTX graphics with NVIDIA's full-stack AI platform in slim, premium Windows PCs, aimed at accelerated creation, immersive gaming and powerful local AI. They are rolling out from ASUS, Dell, HP, Lenovo, Microsoft and MSI.

Research

⭐⭐⭐ [Research] Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline

NVIDIA and IIT Madras' AI4Bharat lab introduced Nemotron-pretraining-Indic-v1, roughly 2.8 trillion tokens across the 22 languages listed in India's Constitution, releasing soon alongside the open, GPU-accelerated data generation pipeline used to create it. The corpus holds 2,601.0 billion native-script tokens and 224.6 billion romanized tokens across 34 language-script configurations, produced through workflows that combine document-level translation, quality assessment, and targeted repair, with checks to preserve equations, tables, and code, and every record traceable to a curated source document and validated against its input. The pipeline shares the full generation recipe — prompts, LLM workflows, and quality checks — so developers can adapt it to their own source material. Continued pretraining for Nemotron 3.5 Super VL added 10 billion Indic tokens drawn from a twelve-language subset of the corpus.