Model components / Inference code

ds4

ds4 (DwarfStar 4) is a native local LLM inference engine by antirez (Salvatore Sanfilippo), optimized first for DeepSeek V4 Flash and deliberately narrow: it serves one model at a time. The self-contained C/CUDA/Objective-C engine builds on llama.cpp/GGML, adding model loading, prompt rendering, tool calling, and KV state streaming to RAM and SSD. It ships an HTTP server with OpenAI-compatible endpoints and a coding agent.

The project has no tagged releases and is built from source, so this entry is verified against the repository head rather than a release. Verified 2026-08-09 via GitHub and the LICENSE body.

Openness

5 high confidence

5.0

license: MIT(OSI)
source: public
core-gated: ungated

MIT-licensed native inference engine; full C/CUDA/Obj-C source public, no proprietary tier.

https://github.com/antirez/ds4/blob/main/LICENSE recorded 2026-08-09
MIT License, Copyright (c) 2026 The ds4.c authors, Copyright (c) 2023-2026 The ggml authors
https://github.com/antirez/ds4 recorded 2026-08-09
repo license metadata reports {"spdxId":"MIT","name":"MIT License"}; single public repository - a full-text read of the page finds no commercial edition, enterprise directory or hosted-service offering, the only "pricing" strings being GitHub's own site navigation, and the project links no separate product site.

Adoption

3 low confidence

3.0

21,258 GitHub stars on antirez/ds4, which falls in the >10K stars band, level 3 on the stars fallback scale; that scale caps at 3 because a star is not a use. No download, install or customer figure is published for this product, so stars remain the only honest signal and the level is directional.

https://api.github.com/repos/antirez/ds4 recorded 2026-08-12
stargazers_count = 21,258 for antirez/ds4

Capability

3 medium confidence

3.0

A specialized single-model-family engine: Metal, CUDA and ROCm backends, DeepSeek V4 Flash/PRO specialization, an OpenAI-compatible HTTP server, RAM and SSD KV streaming, and an integrated coding agent. It has neither the model breadth of llama.cpp nor the datacenter throughput of the frontier serving engines, which places it below both. No MLPerf submission exists for it.

https://github.com/antirez/ds4 recorded 2026-08-09
README states the engine is "deliberately narrow, not a general GGUF runner. Model loading, prompt rendering, tool calls, KV state, the HTTP server, and the coding agent are built and tested together"; confirms Metal/CUDA/ROCm backends and DeepSeek 4 Flash/PRO focus

Verified 2026-08-09