# antirez/ds4

DeepSeek 4 Flash and PRO local inference engine for Metal, CUDA and ROCm

Repository: https://github.com/antirez/ds4
Canonical: https://ross.abutalabs.com/products/ds4
Language: C
License: MIT
License Family: permissive
Last push: 2026-08-23T13:41:58+00:00

## Health v2 (maintenance only)
Score: 58/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 99, release rhythm 35, longevity 8
- inputs: {"age_days": 119, "days_push": 10, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, young
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 21807, forks 2017 (observed 2026-08-28T04:11:32.065727+00:00)

## What it is
DwarfStar (ds4) is a self-contained native LLM inference engine written in C, optimized for DeepSeek V4 Flash and PRO and GLM 5.2, with support for Metal, CUDA, and ROCm backends. It includes model loading, prompt rendering, tool calls, KV caching, an HTTP server, and a coding agent, plus tools for GGUF quantization and benchmarking.

## Use cases
- run DeepSeek V4 Flash locally on a MacBook with Metal
- serve a multi-user LLM server on old CUDA cards with micro batching
- run large models on machines without enough RAM using SSD streaming
- use tensor parallelism across two Macs with RDMA
- run inference on Strix Halo systems with ROCm
- quantize models to GGUF with imatrix and evaluate quality and speed

## When to avoid
- you need a general-purpose GGUF runner for arbitrary models
- you prefer a mature multi-model framework like llama.cpp or vLLM
- you are uncomfortable with AI-assisted developed code

## Facets
- artifact type: application
- maturity: active
- function: llm-inference, http-server, gpu-computing
- domain: large-language-models, artificial-intelligence, self-hosted
- platform: cli
- tags: metal, cuda, rocm, deepseek, gguf, local-inference, tensor-parallelism, ssd-streaming, macos, linux, gpu, web-server

## Member repositories
- antirez/ds4 (main) score 58

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:11:32.065727+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T16:57:47.670860+00:00, confidence not recorded.
  - readme: https://github.com/antirez/ds4 (fetched 2026-08-28T04:11:32.065727+00:00, sha d5943ebf8f53)
- Data as of 2026-08-30T08:39:29.467469+00:00.