google-deepmind/simply

Minimal and scalable research codebase in JAX, designed for rapid iteration on frontier research in LLM and other autoregressive models.

What it solves

Simply is designed to accelerate frontier LLM research by providing a minimal, scalable codebase that is easy for both humans and AI agents to modify and iterate upon. It reduces the time required to implement and test new ideas—such as new optimizers, training losses, or RL algorithms—by removing complex abstractions and heavy dependencies.

How it works

Built with JAX, the project provides a self-contained environment where researchers can rapidly prototype. It includes a variety of tools for training, evaluation, and serving (via gRPC), and supports running experiments on local machines, Google Cloud TPUs, and GKE with XPK. It also features a built-in agent harness and the Amplio harness for long-horizon autonomous research, allowing AI agents to read the code, propose changes, and run experiments autonomously.

Who it’s for

AI researchers and developers who want a minimal codebase for LLM research, as well as those developing AI agents capable of performing automated research tasks.

Highlights

  • Minimalist Design: Uses JAX with minimal abstractions to ensure the code is easy to read and hack.
  • Agent-Ready: Specifically designed as an environment for AI agents to autonomously iterate on research experiments.
  • Scalable Infrastructure: Supports TPU and GKE integration for high-performance training.
  • Sourced Model Zoo: Includes self-contained model plugins for architectures like GLM-5, Kimi K3, and Qwen3.8.
  • Research Benchmarking: Provides a suite of 11 scored research tasks to benchmark the ability of AI agents to conduct research.
  • Amplio Harness: A Go-based agent harness for long-horizon research with persistence and a web UI.

Related

  • Project
    ModelEngine-Group/app-platform

    AppPlatform is a low‑code, Docker‑based platform that lets you visually design, run and share AI applications (chatbots, RAG pipelines, agents). It combines a Java‑plugin backend (FIT + Waterflow) with a React + ELSA front‑end, supports multiple LLMs, and stores reusable app templates. Quick start: copy `.env`, run `docker/deploy.sh`, then open `http://localhost:8001`.

  • Project
    OleksandrChekhovskyi/hax

    hax is a lightweight, terminal‑native AI coding assistant written in C. It talks to cloud or local LLM providers (OpenAI, Anthropic, llama.cpp, Ollama, etc.) via a simple REPL or one‑shot mode, streams Markdown‑formatted answers, and keeps a transparent transcript. Install via Homebrew, AUR, pre‑built binaries, or compile from source; it’s aimed at developers who want a fast, low‑resource, fully auditable AI tool that lives entirely in the terminal.

  • Project
    google/tunix

    A JAX-based post-training library for LLMs that provides scalable support for supervised fine-tuning and reinforcement learning on TPUs.

  • Project
    numtide/llm-agents.nix

    A Nix flake that packages dozens of AI coding agents (terminal tools, desktop apps, and multi‑model orchestrators) for easy installation via `nix run …#<agent>`.

  • Project
    smallnest/pigo

    pigo is a Go‑based CLI AI coding assistant (a clone of the pi agent) that can read/write files, run shell commands, fetch web content, and interact with many LLM providers. It works in headless script mode or an interactive REPL, supports tool‑level allow/deny lists, project trust, prompt templates, skills, plugins, self‑update, and provides a Go SDK for embedding.