changelog github pypi reports

Run any model your machine can handle.

A model-agnostic agentic runtime. See what your computer can actually run — and run it in one command. Local, or a frontier model with your own key. The intelligence is in the harness, not the weights.

# install · macOS / Linux

curl -fsSL https://topk1.com/forge/install.sh | sh

# install · Windows  (PowerShell / cmd)

pip install blueshark-forge
  • Model-agnostic — Ollama, llama.cpp, MLX, or OpenAI / Anthropic with your key.
  • Fits your hardware — laptop, Apple Silicon, or a multi-GPU datacenter.
  • Local & private — runs on your machine, files never leave. stdlib-only.
forge — what can I run?
forge models
Apple M-series · 48GB · GPU / Metal
→ runs well up to ~30B
phi-22.7Bruns well
qwen2.5-coder7Bruns well
llama3.18Bruns well
gemma227Bruns well
sarvam-30b 30Bruns well
# …466 more via `forge models --all`
forge models use phi-2
ready → forge run "fix the failing test"