# DwarfStar 4 (ds4)

> A narrow C inference engine that runs frontier open-weight models locally on high-memory Mac, CUDA and ROCm machines.

DwarfStar 4 (ds4) is a local inference engine created by Salvatore Sanfilippo (antirez), the creator of Redis. It runs a small set of open-weight model families, including DeepSeek V4 and V4.1 Flash, GLM 5.x and Qwen3.8 Flash Next, on high-memory Apple Silicon, NVIDIA CUDA and AMD ROCm machines. The dwarfstar.sh site is a community-maintained resource with docs, hardware guidance and benchmarks around the MIT-licensed engine.

## What It Is

ds4 is a C inference engine that is deliberately narrow rather than a generic GGUF runner. It targets project-specific GGUF layouts that are validated end to end against official model outputs. Models use asymmetric 2-bit quantization on the routed experts while keeping critical shared paths precise, which is how the supported builds fit their target machines.

## How the Stack Works

One engine exposes three interfaces: `./ds4` for interactive chat, `./ds4-server` for local APIs, and `./ds4-agent` for persistent coding sessions. The KV cache can be saved to SSD and resumed by prompt hash, so restarts do not require a full re-prefill. The site lists SSD streaming, tensor parallelism, session batching, DSPARK + MTP and vision input among the capabilities.

## Agent and API Connectivity

ds4-server speaks OpenAI-style and Anthropic-style APIs, with endpoints such as /v1/chat/completions, /v1/messages and /v1/responses. The docs describe connecting OpenCode, Claude Code, Codex CLI and Pi to the local server via a base URL.

## Setup Path

Users clone the repository, download a project GGUF with download_model.sh, and build for their backend (for example make for macOS Metal, or make cuda-spark for DGX Spark). Hardware classes listed include Apple Silicon Macs with 64 GB+, NVIDIA DGX Spark or generic CUDA Linux boxes, and AMD Strix Halo systems. The site publishes benchmark rows, such as 790.2 t/s prefill and 39.4 t/s generation for q2 at 2,048 tokens on an M5 Max with 128 GB.

## Project Approach

The About page says upstream treats the project as a working template that users adapt with coding agents, and that there are deliberately no GitHub releases or tags. Upstream credits llama.cpp and GGML for kernels and quantization formats.

## Features
- Asymmetric 2-bit quantization of routed experts
- KV cache persisted to SSD and resumed by prompt hash
- Interactive CLI (ds4)
- Local server with OpenAI and Anthropic-style APIs (ds4-server)
- Native persistent coding agent (ds4-agent)
- SSD streaming
- Tensor parallelism
- Session batching
- DSPARK + MTP speculative decoding
- Vision input
- Metal, CUDA and ROCm backends

## Integrations
OpenCode, Claude Code, Codex CLI, Pi, llama.cpp, GGML

## Platforms
MACOS, LINUX, CLI, API

## Pricing
Open Source

## Links
- Website: https://dwarfstar.sh
- Documentation: https://dwarfstar.sh/docs/quickstart/
- Repository: https://github.com/antirez/ds4
- EveryDev.ai: https://www.everydev.ai/tools/dwarfstar-4