AI Daily — 2026-04-27

English 中文

Microsoft Unveils TRELLIS.2: Open-Source 4B 3D Generator · Local browser AI agent runs offline wi...

Covering 33 AI news items

🔥 Top Stories

1. Microsoft Unveils TRELLIS.2: Open-Source 4B 3D Generator

TRELLIS.2 introduces a 4B-parameter model for image-to-3D generation that uses a novel O-Voxel sparse voxel structure to produce high-fidelity assets with full PBR materials. It supports 16x spatial compression via native 3D VAEs and can generate assets up to 1536^3. The project is open-source, with links to the paper, GitHub code, and a HuggingFace live demo. Source-reddit

2. Local browser AI agent runs offline with WebGPU

Introduces a completely local browser-based AI agent powered by Gemma 4 E2B and WebGPU. The agent uses native tool calling to search history, read and summarize pages, and manage tabs, all 100% offline with no servers. It also enables features such as HLS playback inside the browser. Source-x

3. Xiaomi opens MiMo-V2.5 under MIT License for commercial use

Xiaomi released MiMo-V2.5 as an open-source MIT-licensed model, enabling commercial deployment, continued training, and fine-tuning without extra authorization. The release includes MiMo-V2.5-Pro for complex agent and coding tasks (top-ranked on GDPVal-AA and ClawEval) and MiMo-V2.5, a native multimodal model with strong agent capabilities, both featuring a 1M-token context window, with weights on HuggingFace and details on the Mimo Blog (mimo.xiaomi.com). Source-x

📰 Featured

Open Source

DeepSeek AI v4 paper praised for open, rigorous research — The paper is lauded for mathematical rigor and open release, highlighting that months or years of work were released openly for free to benefit the wider AI community. Source-x
Hermes Agent Repo Surpasses Claude Code Repository — Signals growing momentum in open-source AI tooling and competitive dynamics among AI code repositories. Source-x

AI Safety

Normative Conflicts Increase Jailbreaks in Large Reasoning Models — An arXiv study shows normative conflicts raise jailbreak risk in large reasoning models, arguing current alignment is shallow and safety gaps persist. Source-x
Creator Builds With GPT-5.5, Highlights Claude’s Flaws — A project built with GPT-5.5 contrasts against Claude’s persistent issues, teasing a compilation of observed flaws and something called “clawd rip.” Source-x

LLMs / Benchmark

Nemotron-3 Nano Wins 4B Benchmark Across Finance, Reasoning, Code — In a head-to-head across 39 tasks, Nemotron-3 Nano leads 4B-class models, with strong scores in finance and reasoning and solid code performance. Source-reddit
Kimi K2.6 Tops OpenRouter Weekly LLM Leaderboard — Kimi K2.6 rises to No. 1 on OpenRouter’s weekly leaderboard, reflecting rapid ecosystem progress and continued development. Source-x

Hardware / Inference

Skymizer Ultra-Large LLM Inference on a Single PCIe Card — A single PCIe card with six HTX301 chips and 384 GB memory reportedly runs 700B-parameter LLM inference locally at ~240W per card, reducing reliance on high-VRAM GPUs; more details expected at Computex. Source-reddit

⚡ Quick Bites

UniT Enables Unified Physical Language for Humanoid Transfer — Introduces a unified physical language for transferring tasks to humanoid agents. Source-huggingface
Luce DFlash Delivers 2x Qwen3.6-27B Throughput on RTX 3090 — Achieves up to 2x throughput for Qwen3.6-27B on an RTX 3090. Source-reddit
Open-source Cua Enables AI Agents to Control Full Desktops — Enables AI agents to interact with and control full desktop environments. Source-github
China Blocks Meta’s Manus Acquisition by Foreign Investor — China blocks Meta’s Manus acquisition by a foreign investor. Source-reddit
2x 5060 Ti: Better Qwen 3.6 27B/35B Configs? — Community discussion on improved Qwen 3.6 configurations. Source-reddit
OpenAI Privacy Filter Runs On-Device via ExecuTorch — Privacy filter can run on-device using ExecuTorch. Source-reddit
GitHub Copilot Moves to Usage-Based Billing on June 1 — Copilot shifts to usage-based billing starting June 1. Source-x
Symphony: Open-source Codex agent orchestrator for issues — Symphony provides an open-source Codex-like agent orchestrator for issue handling. Source-x
Agentic World Modeling: Foundations, Capabilities, Laws, and Beyond — New perspective on agentic world modeling foundations and capabilities. Source-huggingface
gpt-realtime-1.5 Enables Voice-Controlled Interactive Apps — Adds voice control to interactive apps via gpt-realtime-1.5. Source-x
AI Firms Must Benchmark to Benefit from Model Progress — Calls for benchmarking as a driver for benefiting from model progress. Source-x
Semantic Progress Function for Video Analysis and Generation — Introduces a semantic progress function for video analysis/generation. Source-huggingface
mattpocock/skills: Agent Skills for Real Engineers — Repository sharing agent skills for engineers. Source-github
Beads: Dolt-backed memory graph for AI coding agents — Beads memory graph for AI coding agents. Source-github
OpenClaw Launches Personal AI Assistant On Your Own Devices — Personal AI assistant that runs on user devices. Source-github
To 16GB VRAM users, plug in your old GPU — Practical guidance for users with 16GB VRAM to reuse older GPUs. Source-reddit
GBNF tweak speeds Qwen3.6 35B-A3B and 27B — Grammar/numeric tweaking speeds for Qwen3.6 models. Source-reddit
End-to-End Fine-Tuning Tutorial for Multimodal Models — Tutorial covering end-to-end fine-tuning of multimodal models. Source-reddit
Grok 3 Open-Source Delay Sparks Skepticism — Opensource delay for Grok 3 sparks skepticism. Source-reddit
Compilation Shows Claude’s Issues Over the Years — A thread compiling historical issues with Claude. Source-x
Qwen 3.6 27B on Strix Halo 128GB: experiences? — User experiences with Qwen 3.6 27B on Strix Halo 128GB. Source-reddit
Local LLMs on 56GB VRAM: Best model with llamacpp — Discussion on running local LLMs on 56GB VRAM with llamacpp. Source-reddit
Multi-Agent Collaboration Highlighted on Twitter/X — Highlighting multi-agent collaboration threads. Source-x

Generated by AI News Agent | 2026-04-27

━━━━━━ End of Template ━━━━━━