
Google's Antigravity SDK Lets Gemma 4 Run Agent Workflows Without the Cloud
Google's Antigravity SDK now runs agentic workflows entirely offline using Gemma 4 26B and LiteRT, with hybrid cloud-local orchestration for privacy-sensitive code.
A focused feed of models, agents, research and open-source releases for people building with AI.

Google's Antigravity SDK now runs agentic workflows entirely offline using Gemma 4 26B and LiteRT, with hybrid cloud-local orchestration for privacy-sensitive code.

Black Forest Labs releases a 7B open-weight world action model that tops RoboLab, cutting parameters by 56% and running up to 3.95x faster.

NVIDIA's 100M-parameter Nemotron 3 Diarization tracks up to 8 speakers in overlapping conversations, topping Voice Arena's benchmark with 14.72% DER.

OpenHarness ties every coding agent to persistent tmux sessions across your machines, adds domain-specific specialists for CAD, PCB and robotics, and ships an open-hardware desk device.

CopilotKit released OpenMuse, an MIT-licensed personal agent with its own browser, terminal, and file system that runs on iOS, Android, and web.

A 53M-parameter open-vocabulary scene-graph model runs at 20ms per frame, understands 19,103 relation phrases, and needs no object labels.

Altworld's 27B fine-tune of Qwen3.8 tops GPT-6 Astra on everyday message writing and reads as human 26 points more often.

Yandex open-sourced an 80B mixture-of-experts model with 3B active parameters, 256K context, and a hybrid KDA-attention architecture trained from scratch.

Unreal Labs open-sourced a Go-based async agent harness that reportedly runs Terminal-Bench 4.0 at 39 percent lower cost than Codex plus Astra.

White Circle open-sourced Halo, a distributed training framework that keeps models in native HuggingFace format while hitting up to 2.8x TRL throughput.

A 3,500-star open source rules pack forces AI coding agents to strip out generic hero sections, invented metrics, and useless comments before shipping.

NetEase Youdao's open-source R2T2 streams speech to text with 200-600ms latency and never rewrites words it already emitted.

Xiaomi's new 1.02T parameter MoE tops the Artificial Analysis open weights leaderboard while costing pennies per task through aggressive caching and sparse activation.

An independent project rebuilds TypeSafe's closed Jev decision service using open 4B models on a single RTX 3090, with published benchmarks.

An independent tinkerer publishes ComfyUI workflows, custom nodes, and benchmarks that push MiniMax H3 video generation to 3.12x speedup across four RTX 3090s.

Jared Palmer's open-source Kev family scales a Jev-style decision model up to 8B parameters using Qwen3, LoRA, and a pointer head.

Z.ai just open-sourced ZCode, a full coding agent workbench spanning desktop, web, and terminal, tuned for GLM-5.3 but tool-agnostic.

Bespoke Labs open-sources Nimble, a 9B model that makes typed decisions without generating tokens, using contrastive pairs for training.

OpenHuman is an open source, local-first agent harness that gives your AI a persistent memory of your life and orchestrates fleets of agents on your machine.

Laya is an open-source, non-autoregressive decision engine that answers typed questions across 100+ languages in a single 33ms forward pass.

Alibaba open sources a 7B image model with native RGBA output, 10 reference image editing, and unified generation plus editing capabilities.
A native MLX port of the Laya typed decision model runs 7-14 ms structured classifications locally on Apple Silicon, no text generation required.

Tencent Hunyuan open-sources a 1.5B speech foundation model that handles TTS, editing, enhancement, and separation through plain-language instructions.

Tencent Cloud open-sourced Octop, a single-process, self-hosted AI assistant that gives every user their own team of MBTI-flavored agents across web, CLI, and chat apps.
No stories match these filters.