Model components / Inference code

ds4

Native local LLM inference engine by antirez (Salvatore Sanfilippo), optimized first for DeepSeek V4 Flash (with DeepSeek V4 PRO support on very-high-memory machines). Self-contained C/CUDA/Objective-C engine built on llama.cpp/GGML with DeepSeek-specific loading, prompt rendering, tool calling, and KV state streaming (RAM and on-disk SSD). Ships an HTTP server with OpenAI-compatible endpoints and an integrated coding agent, across Metal (Apple Silicon), NVIDIA CUDA/DGX Spark, and AMD Strix Halo (ROCm) backends.

MIT-licensed (LICENSE body, (c) 2026 The ds4.c authors + ggml authors). Created 6 May 2026, ~332 commits, last push 17 Jun 2026; ~14.6k stars, no tagged releases (built from source). Deliberately narrow 'one model at a time' design. Verified 2026-06-19.

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 2026-06-19
MIT License text, (c) 2026 The ds4.c authors; (c) 2023-2026 The ggml authors

Adoption

2 low confidence

2.0

~14.6k GitHub stars six weeks after launch, but driven heavily by antirez's profile; no package distribution and no install/usage count, so treated as a low-confidence star-proxy pending warehouse signal.

https://github.com/antirez/ds4 recorded 2026-06-19
~14.6k stars, 332 commits, 0 releases, created May 2026

Capability

3 medium confidence

3.0

Specialized single-model-family engine rather than the breadth of llama.cpp (C4) or the datacenter-throughput frontier (C5); scored 3 for a focused, capable local runtime.

https://github.com/antirez/ds4 recorded 2026-06-19
README documents Metal/CUDA/ROCm backends, OpenAI-compatible server, SSD KV streaming, integrated agent

Unchanged since 2026-07-30 (last edited, not re-checked)