ivo's home

Hugging Face's work on the PyTorch 2.15 MPS backend brings FP8 support, native grouped matmuls, and dedicated Metal kernels to Apple Silicon Macs running macOS 15. Mixture-of-experts models see decoding speedups up to 6.5 times faster on hardware like the M3 Ultra, while memory peaks for video VAE decoding drop from gigabytes of overhead to manageable footprints. The release replaces various MPSGraph bottlenecks with direct Metal execution for single-row linear layers, softmax, and common linear algebra operations.

It is worth adopting if you are actively running local inference or fine-tuning models on a modern Mac and hitting previous memory or fallback walls. However, the catch is the strict hardware floor requiring macOS 15 and Python 3.11, meaning older setups get left behind without an upgrade path. Skip this if you are strictly bound to CUDA infrastructure, as parity with NVIDIA hardware is still incomplete.

more: https://huggingface.co/blog/mps-pytorch-2-15

my other side: ivo.blogspot.com

DigUp is an open-source desktop app for macOS, Windows, and Linux that indexes local files for semantic search using Google DeepMind's EmbeddingGemma 2 model. It runs entirely on your hardware, processing text, images, audio, and video into a 740M parameter embedding space so you can query files using natural language. The initial run downloads an 865 MB model and indexes your selected directories locally without sending data to external servers.

It is worth installing if you regularly hunt through unindexed mixed-media directories and want something faster than grep or Spotlight for conceptual queries. The catch is that it requires a modern machine to index smoothly and relies on external players to jump directly to specific timestamps in media files. Skip it if you only need exact-match text searches or refuse to run local ML models in the background.

more: https://github.com/ARahim3/DigUp

my other side: ivo.blogspot.com

CodeAgent is an open-source, terminal-based AI programming assistant written in Python 3.11 with extensive Chinese commenting. It handles natural language instructions by reading files, executing commands, searching codebases, and running multi-model loops via OpenAI-compatible or Anthropic endpoints. The tool features a built-in session store, hierarchical context compression, permission panels, and a self-learning skill system stored locally under ~/.codeAgent.

It is worth checking out if you want a local, hackable alternative to proprietary CLI tools with built-in DeepSeek support and clear module architecture. However, Windows is the primary development target with Linux support trailing, while macOS remains largely unverified. Skip this if you need out-of-the-box cross-platform stability without configuring sandbox environments or API keys manually.

more: https://github.com/Harvil1/codeAgent

my other side: ivo.blogspot.com

StartLux-Decision provides open models in 4B, 9B, and 27B parameter sizes that output typed probabilities for structured questions rather than open text. The repository includes an HTTP server and inference engine supporting GGUF and MLX formats, with variants that conditionally generate reasoning traces when confidence falls below 0.6. It handles text, JSON, and image inputs in a single forward pass, making it suitable for programmatic control tasks like game bots and video frame analysis.

If you need structured classification or control loops without parsing free-form LLM output, it is worth testing locally via the provided MLX or GGUF weights. However, the specialized API and decision-focused architecture mean it will not replace a general-purpose chat model or handle open-ended text generation. Skip it unless you are actively building deterministic pipelines that require explicit probability scores for discrete options.

more: https://github.com/StartLuxLabs/StartLux-Decision

my other side: ivo.blogspot.com

The @receptron/laya npm package runs Convai Innovations' Laya decision model directly in Node.js and TypeScript via ONNX Runtime without needing Python or PyTorch. It takes a JSON state and typed questions, returning calibrated probabilities for choices, scores, and yes-no statements in a single forward pass taking roughly 140 milliseconds per batch on Apple silicon. The 1.7 GB model weights download automatically on first use, and the TypeScript types match the Python implementation to four decimal places.

It is worth testing if you need structured classification or confidence scoring on text inputs without spinning up an LLM API. The catch is strict input limits: states are truncated at 512 tokens and choice questions top out around 20 options. Skip this if you need generative text or inputs longer than a short support ticket.

more: https://github.com/receptron/laya

my other side: ivo.blogspot.com

Talorys is an open-source personal AI assistant that deploys entirely inside your own Cloudflare account using a single npm command. It packages a chat interface, SQLite-backed durable memory, task management, and scheduled alarms into a private Pages and Worker setup without external servers or databases. It runs on Cloudflare free-tier services using Workers AI and the Agents SDK for streaming responses and tool calling.

If you already use Cloudflare and want a private workspace for tasks and notes without paying for SaaS subscriptions, it takes five minutes to spin up. It is strictly single-user with no multi-tenant features, and relies on Cloudflare's free limits and infrastructure. You will need Node 22 and an active Cloudflare login to complete the initial setup.

more: https://github.com/rociiu/talorys

my other side: ivo.blogspot.com

Allen Institute's HiRO-ACEv2 is an open-source machine learning system that downscales coarse 100 km climate data to 3 km resolution. It pairs a coarse atmospheric emulator, ACE, with a fine-scale diffusion model, HiROv2, to simulate rare rain, wind, and pressure events. The full pipeline generates global high-resolution output at roughly one-three-hundredth the compute cost of traditional numerical models like X-SHiELD.

The models and tutorial notebooks are available on Hugging Face if you want to inspect extreme weather generation or run regional downscaling yourself. The major catch is that HiROv2 was trained exclusively on X-SHiELD outputs, meaning it currently only works with that specific ACE baseline and cannot yet ingest historical datasets like ERA5 or projected future climates.

more: https://huggingface.co/blog/allenai/hiro-acev2

my other side: ivo.blogspot.com

AQuA is a framework for diagnosing agent behavior by analyzing production telemetry via Cloud Trace, Logging, or BigQuery. It functions as an ambient quality agent that monitors live sessions to identify silent failure patterns and attribute them to specific code revisions or upstream data sources. The system operates outside the request path and generates actionable, code-anchored insights for human review.

This is worth implementing if you are struggling to manage the drift between your static evaluation datasets and live user traffic. It is not a silver bullet for deployment, as it requires you to write domain-specific goal files and deterministic metrics to be effective. Expect to spend significant time tuning the evaluation prompts before the insights become truly reliable.

more: https://developers.googleblog.com/the-outer-loop-insights-first-an-ambient-quality-agent-that-diagnoses-your-production-agent/

my other side: ivo.blogspot.com

Agent Lightning v1.0 is a 3,500-line open-source framework from Microsoft Research for applying reinforcement learning to existing agent harnesses. Instead of forcing you to rewrite your agent's execution loop inside the training framework, it sits between your agent and the LLM API as an OpenAI-compatible proxy. It records model calls and handles collocated asynchronous training on a shared pool of GPUs.

It is worth a look if you are already maintaining a custom harness like OpenHands or mini-SWE-agent and want to tune it without rewriting your stack. The catch is that splitting rollouts into variable training samples introduces complex data alignment challenges, and it assumes you have the compute infrastructure to run heavy concurrent agent loops. Skip it if you are looking for an out-of-the-box training environment with zero systems overhead.

more: https://www.microsoft.com/en-us/research/blog/agent-lightning-v1-0-a-3500-line-lightweight-agentic-rl-framework-for-training-agents-with-real-harnesses/

my other side: ivo.blogspot.com

Pinvou Agent is an open-source desktop workspace that combines document chat, vector knowledge bases, and design generation into a single Electron application. It supports local models via vLLM alongside any OpenAI-compatible endpoint, and lets you hook up CLI tools and MCP servers. The codebase also includes ACP support for plugging in coding agents like Claude Code or Codex against a local project directory.

If you are tired of juggling CLI windows and browser tabs for different AI tasks, this consolidates them neatly under ~/.pinvou3/ with persistent sessions and local voice input. However, it is a heavy desktop wrapper that still requires managing your own model backends and API keys for cloud endpoints. Skip it if you prefer keeping your workflow strictly inside a terminal or your existing text editor.

more: https://github.com/Pinvou/pinvou-agent

my other side: ivo.blogspot.com