ivo's home

Yuvraj Sharma and Abubakar Abid describe using an agentic HuggingChat tool called ML-intern to turn large models into tiny, CPU-runnable custom variants. For sixteen dollars of compute, the system takes an initial prompt and dataset, handles training and evaluation on Hugging Face hardware, and outputs models like a 0.8B Qwen-Image rewriter that replaces a 20GB teacher model. The process relies on detailed specification prompts that enforce baseline testing and strict budget limits.

It is worth trying if you routinely need task-specific models that fit on local hardware without paying API fees. The catch is that the output quality depends entirely on writing structured prompts up to two thousand words long, including verified facts and smoke tests. If you prefer turnkey tooling or do not want to curate training datasets, skip this approach.

more: https://huggingface.co/blog/building-with-ml-intern

my other side: ivo.blogspot.com

Magpie is a cross-platform menu bar app and local proxy that unifies dozens of AI coding agents behind standard wire APIs. It translates between OpenAI, Anthropic, and Gemini formats on localhost while managing quotas across multiple accounts and subscription logins. Instead of editing separate JSON config files for every tool, you pick a model from a list and the gateway handles failover automatically when limits hit.

It is worth installing if you juggle multiple CLI agents and paid accounts, saving you from manually rotating API keys when quotas run out mid-task. The catch is that routing traffic through an extra local proxy adds another moving part to debug when requests fail or tokens do not match up. Skip it if you only use a single official client and standard API keys without hitting limits.

more: https://github.com/yetone/magpie

my other side: ivo.blogspot.com

Liquid AI has released d1-3B and d1-omni-600M, two decision models designed to perform classification and extraction tasks in a single forward pass. These models move away from generative token output, delivering responses on edge hardware like the Jetson AGX Orin in under 50 milliseconds. The 3B parameter model handles text, vision, and audio, while the smaller 600M version targets footprint-constrained environments.

If you are building structured data pipelines or classification tasks on the edge, this is worth testing to avoid the overhead of a full generative LLM. However, be aware that you must use trust_remote_code for the custom inference architecture, which introduces a dependency on their specific implementation. It is not a general-purpose conversational agent, so look elsewhere if you need chat capabilities.

more: https://huggingface.co/blog/LiquidAI/open-d1

my other side: ivo.blogspot.com

OpenQodex is a command-line tool that automates code review by leveraging local instances of Claude Code or Codex. It identifies uncommitted changes, runs targeted scanners, and produces a local HTML report summarizing findings without requiring external accounts or servers. It requires Node 22, Git, and a Unix-based environment like macOS, Linux, or WSL.

It is a useful utility if you already rely on coding agents and want a structured way to enforce local quality gates before pushing. However, it is not for those looking for a zero-config setup, as it demands manual dependency management and relies heavily on the performance and behavior of the underlying AI reviewers. If you prefer automated CI pipelines over local terminal workflows, this will likely feel redundant.

more: https://github.com/openqodex/openqodex

my other side: ivo.blogspot.com

Google has released EmbeddingGemma 2, a sub-1 billion parameter model that maps text, code, images, video, and audio into a shared 768-dimensional space. Integrated into sentence-transformers v6.1.0, the architecture is modular so you can drop unused modality encoders at load time to scale the footprint from 270M parameters up to 740M. It includes support for task-specific prompts and native dimension truncation down to 128 dimensions.

If you are building a local RAG pipeline or cross-modal search tool that needs to handle mixed media without a heavy inference server, this is worth testing. The catch is that matching modalities still requires strict adherence to prompt names and specific token markers, and sub-1B models will inevitably lag behind larger proprietary endpoints on nuanced retrieval tasks. Skip it if your stack is purely text-based and already tuned around existing dense retrievers.

more: https://developers.googleblog.com/embeddinggemma-2-the-developer-guide/

my other side: ivo.blogspot.com

Agent Memory is a runtime that gives AI agents local, file-based persistence by mapping conversations to a directory of Markdown files indexed by SQLite. It functions as a local cache that agents can query via shell commands, allowing them to recall context across sessions without relying on external databases or cloud APIs. The system treats the filesystem as the source of truth, meaning all data remains readable by standard tools like grep even if the index is deleted.

It is worth your time if you prefer local, transparent, and git-managed data over opaque vector database abstractions. However, it requires you to manage a directory-based workflow and configure individual agent schemas, which adds overhead if you just want a plug-and-play solution. It is not for users who prefer managed services or have no interest in maintaining their own data structure.

more: https://github.com/tigerless-labs/agent-memory

my other side: ivo.blogspot.com

Microsoft and Hugging Face released ThinkingBox via OpenEnv to evaluate AI agent reliability across 507 stateful business workflows. Instead of checking generated text or tool calls, it runs agents in isolated Model Context Protocol sessions and checks the resulting database records. Running 12 LLM models across 121,680 trials, the benchmark measures whether an agent can produce the correct backend state twenty times in a row.

It is worth running if you are building autonomous database workflows and need to move past single-run pass rates. The catch is that most current models suffer drastic reliability drops when forced to repeat tasks consistently, retaining as little as eight percent of their pass@1 score. Skip this if you only need conversational assistants or static text generation.

more: https://huggingface.co/blog/microsoft/thinkingbox

my other side: ivo.blogspot.com

The dsh-our-free-model repository provides a zero-dependency plugin for the dsh runtime that exposes models like DeepSeek V4.1 and Kimi K3. Installation requires no API keys, accounts, or configuration, routing requests directly through the OpenCode Zen gateway. It includes a local OpenAI-compatible forwarding port, a usage dashboard, and automatic hot-reloading.

It is worth installing if you already use dsh and want free access to frontier models without managing credits. However, availability depends entirely on your local network's exit IP, and heavily managed desktop bundles may disable the plugin's auto-update features. Skip it if you require guaranteed uptime or enterprise SLAs.

more: https://github.com/Ebony-Vinyl/dsh-our-free-model

my other side: ivo.blogspot.com