Vannarot Roeung
Recent posts by Vannarot Roeung
AMD Ryzen AI Halo: $3,999 Local AI Workstation
AMD put a 128GB local AI workstation on retail shelves for $3,999. The Ryzen AI Halo runs models up to 200 billion parameters and undercuts Nvidia's DGX Spark by $700.
How to Deploy an MCP Server to Claude and ChatGPT
Take an MCP server from an empty folder to a live integration Claude and ChatGPT can both call, in about 30 minutes. Exact commands with mcp-use and Manufact Cloud.
OfficeCLI: Office Suite Built for AI Agents
OfficeCLI is an open-source binary that lets AI agents read, create, and edit Word, Excel, and PowerPoint files with no Microsoft Office installed.
Tencent Hy3: Open 295B Agentic Coding Model
Tencent open-sourced Hy3, a 295B mixture-of-experts model (21B active, 256K context) under Apache 2.0. It beats GLM-5.2 at roughly half the size and is free on OpenRouter until July 21.
OpenAI gpt-realtime-2.1: Faster Voice AI API
OpenAI released gpt-realtime-2.1 and gpt-realtime-2.1-mini on July 6, two Realtime API voice models that cut p95 latency at least 25% and add a lower-cost tier for speech-to-speech agents.
Gladia CLI: Solaria Speech-to-Text From Your Terminal
Gladia shipped a command-line version of its speech-to-text platform, turning audio transcription into one terminal command with SRT and VTT subtitles, speaker diarization, and 100-plus languages.
Ternlight: 7MB Embedding Model Runs in the Browser
Ternlight packs semantic search into a 5 to 7 MB WebAssembly bundle that runs on the CPU with no API, and it hit the Hacker News front page within a day of launch.
Hugging Face to SageMaker Studio: One-Click Deploy
AWS shipped one-click Deploy and Customize buttons on Hugging Face model pages that drop you straight into SageMaker Studio with the model preloaded.
Claude Code Ports C&C Generals to Mac and iPhone
An AI coding agent, Claude Code running Anthropic's Fable model, engineered a native port of Command and Conquer Generals: Zero Hour to Apple Silicon, iPhone, and iPad.
ResearchStudio-Reel: Turn Papers Into Video and Blogs
Microsoft open-sourced ResearchStudio-Reel, five Claude Code and Codex skills that turn one document into a consistent poster, talk video, and bilingual blog in a single pass.
LongCat-2.0: China's 1.6T Open-Weights Coding Model
Meituan open-sourced LongCat-2.0, a 1.6-trillion-parameter agentic coding model trained entirely on Chinese chips. The weights just landed on Hugging Face, and it beats GPT-5.5 on SWE-bench Pro.
Claude Fable Reviewed sqlite-utils Before Its Release
Simon Willison had Claude Fable review sqlite-utils 4.0 before shipping. The agent caught a silent data-loss bug that human review missed.
Orion4D MetaPrompt: Local Prompt Engineering for ComfyUI
Orion4D MetaPrompt is a free, MIT-licensed ComfyUI suite that builds prompts from text lists and enhances them with local Ollama models, fully offline.
pxpipe Cuts Claude Token Bills 70% by Imaging Context
pxpipe is a free, open-source proxy that cuts Claude Code token bills by 59 to 70 percent by rendering bulky context as images the model reads back with OCR. Version 0.7.1 shipped July 3.
Manufact Launches MCP Cloud for Claude, ChatGPT Apps
Manufact, the startup behind the open-source mcp-use framework, has launched Manufact Cloud, a hosting platform that takes an MCP project from a GitHub push to a live endpoint Claude, ChatGPT, Gemini, and Copilot can call.
Condense Proxy Cuts Claude Code Token Bills up to 70%
Condense, a new context-compression proxy for AI coding agents, launched July 3 2026 with a claim aimed at heavy Claude Code users: up to 70% lower token bills with almost no quality loss.
Leanstral 1.5: Mistral's Open Lean 4 Proof Agent
Mistral's open-weights Leanstral 1.5 is a free Apache-2.0 Lean 4 code agent that beats Opus 4.6 on FLTEval at one-seventh the cost.
NVIDIA Nemotron TwoTower: 2.4x Faster Open Diffusion LLM
NVIDIA's Nemotron-Labs-TwoTower generates text 2.42x faster at 98.7% quality by adding a denoiser tower to a frozen autoregressive backbone, no retraining required.
How to Build Reusable Figma Agent Skills in 2026
Build a reusable Figma agent Skill in minutes, trigger it with a slash, and wire MCP connectors to Slack, Notion, and Asana.
First Open-Source Diffusion ASR Model: How It Works
Interfaze's diffusion-gemma-asr-small is billed as the first open-source diffusion-based speech recognition model, refining random tokens into a transcript instead of decoding left to right.
ZeroLabs: Free Open-Source ElevenLabs Alternative
ZeroLabs stitches six open-source models into one free browser studio for voice cloning, design, cleanup, sound effects, and transcription, running on Hugging Face ZeroGPU.
Runway Agent Skills: Ad Campaigns From One Command
Runway shipped Agent Skills on July 2, one-command actions inside its marketing agent that build an ad campaign, create a commercial, or localize your ads from a single prompt, live across all plans.
Claude-Real-Video: Let Any LLM Actually Watch a Video
claude-real-video is a local CLI that pulls scene-change frames, dedupes repeats, and transcribes audio so Claude, ChatGPT, or Gemini can actually watch a video, not just read its transcript.
ZCode: Zhipu's GLM-5.2 Coding Agent, Chat Triggers
Zhipu has shipped ZCode, an official desktop coding agent built on its open-weight GLM-5.2 model, and it lands with a feature no other coding agent ships today: you can trigger and advance tasks from a chat app.