A focused feed of models, agents, research and open-source releases for people building with AI.
Google released Gemini Omni 1.2 Flash in AI Studio and the Gemini API 30% cost reduction versus Omni 1.1 Flash, with faster generation times Native clip length extended to 15 seconds from previous 10-second cap Conversational edit loop lets you refine shots without regenerating from scratch
Google released Gemini Omni 1.2 Flash , now generally available on the Gemini API and Gemini for business. Native video clip length jumped from 10 to 15 seconds, adjustable in 1-second increments from 3 to 15. API pricing dropped 30% to $0.07 per second for 720p video output.
Perplexity launched Teammates , persistent AI agents that own ongoing work on top of Computer. Each teammate runs in an isolated desktop and browser with independent sessions and cookies. Teammates inherit your existing Projects, skills, memory, and tool permissions from Computer.
Reflection published its safety training and evaluation methodology for Beam, its 501B open-weight model. Beam refuses 89% of harmful requests while answering 82% of safety-adjacent benign ones, topping open-weight peers.
Meta launched nts.meta.com , a free public time service using Network Time Security (RFC 8915). Every NTP packet is authenticated so clients verify time came from Meta and was not tampered with. Servers are stateless: cookie sealing keys derive from a shared master and the Unix epoch day number.
Google DeepMind released EmbeddingGemma 2 , a 740M open multimodal embedding model under Apache 2.0. Unifies text, code, images, video, and audio in one 768-dim vector space, replacing caption-then-embed pipelines.
Parallel MCP 2.0 ships four tools behind one URL: web_search, web_fetch, web_decisions, web_research Free and keyless for search, fetch, and decisions at lower rate limits web_research requires OAuth sign-in or an API key, billed at underlying API rates Decisions returns typed answers:
Google Research published in Nature an AI that detects hypertension from smartwatch PPG and accelerometer signals. Model is called AI-PPG-ACC-HTN and requires no cuff calibration, unlike existing cuffless wearables.
Prism ML released Ternary-Bonsai-2-9B-gguf , a 9B VLM with ternary weights at ~2.42 GB. Compresses roughly 7.5x from FP16 using {-1, 0, +1} weights with FP16 group scales. Inherits Qwen3.5-9B hybrid attention and a 262,144-token context window.
Apache Spark 4.3 ships in Databricks Runtime 19 and open source Real-Time Mode gains stateful ops with sub-150ms P99 latency across three benchmarks New PipelinedShuffleDependency lets producer and consumer stages run concurrently Auto CDC adds SCD Type 2 for history-preserving pipelines via
Mistral expands Studio with managed agents, apps, and workflow improvements powered by Mistral Large 4 (Le Chonk). Managed Agents (preview) include skills, sandboxes, subagents, and hooks for custom execution logic.
MiniMax released Music 3 , an open-weight text-to-music model generating full 5-minute songs at 32 kHz stereo. Architecture pairs an 8B Qwen3-based global LLM with a 0.6B local LLM and Flow-VAE synthesis. Native ComfyUI support via the Comfy-Org repack , runs on ~22 GB VRAM with offloading.
webAI released TwIL-8B , an 8.79B reasoning model post-trained from IBM's Granite 4.2 8B. Scores 98.2 on MATH-500 and 81.6 on LiveCodeBench, topping its weight class. Reaches 88 to 93% of frontier model scores on contest math, code and instruction following.
Switch is an open-source platform putting AI teammates into Slack, Teams, Discord, Telegram and Mattermost channels. Channels become shared rooms where people and multiple agents work from the same context and history.
NeuralCrop is a differentiable hybrid of LPJmL and neural networks for crop yield projection. Trained first to emulate LPJmL, then fine-tuned on flux-tower and yield data end-to-end. Improves yield correlation by 80.6% for European wheat and 25.6% for US maize versus baseline.
Comfy-Org repackaged RelayVSR as int8 ComfyUI-ready safetensors for 2× and 4× video upscaling Uses sparse Wan2.2 5B keyframes every 15 frames plus a lightweight Flash decoder per frame Hits 12.59 FPS at 4096×3072 with the LightX2V-optimized variant on a single GPU Native 2× variant targets
Tiny corp launched the tinybox red whale at $50,000 with 14x AMD R9700 GPUs. Total GPU memory hits 448 GB GDDR6 with roughly 8.96 TB/s aggregate bandwidth. Ships in the tall dual-PSU chassis also used for the upcoming 4x MI300X build.
Baidu released ERNIE 5.5 Flash Base , a 280B sparse MoE with 14B active parameters per token Pre-trained on 33.5T tokens with 128K training context, extended to 1M tokens for fine-tuning and inference Hybrid attention: 37 Heavily Compressed Attention layers plus 6 MLA+DSA layers cut KV cache
Tencent released Hunyuan Hy4 , scoring 35 on Artificial Analysis Intelligence Index v5. 770B parameter MoE with 1M token context, vision input, and computer-use via GUI and terminal. Priced at $0.834/$2.501 per 1M input/output tokens, roughly 10x cheaper than Grok 4.7 at same score.
rea-video is an Apache-2.0 agent skill that reverse-engineers short videos shot by shot. Classifies each shot as live action, AI-generated, digital human, 3D, code animation or screen recording. Exports SRT/ASS subtitles, caption styling, cut timing, BPM and voiceover type per clip.
Antigravity now renders .ipynb files natively, including code, Markdown, tables, images, and outputs Feature is view-only: no built-in cell execution, extensions still needed to run code Ships alongside remote cloud agents and Gemini Live voice mode , both rolling out gradually Removes the need to
SGPnP plugs pretrained score-based diffusion models directly into PnP solvers without reverse diffusion sampling. Injecting Gaussian noise into the denoiser input helps iterates escape strict saddle points in ill-posed problems.
Krea2_Asian_Character LoRA trends on Hugging Face with roughly 864K downloads on the model page . Trained on Krea 2 Raw using AI-Toolkit, targeting Japanese character styles with varied framing. Works on both Raw and Turbo checkpoints, following Krea's train-on-Raw, run-on-Turbo workflow.
Benji Taylor released benjitaylor/skills , a public collection of Agent Skills, now at 2.1k stars First skill is Stageshow: React kit for faux product UI stills, cursor demos, bento grids Install with npx skills add benjitaylor/skills --skill stageshow into any Agent Skills-compatible agent Ships
No stories match these filters.
