ds4 is MIT-licensed and provides model-specific quants, local APIs, persistent KV caches and multiple backends.
DwarfStar builds a narrow local runtime for frontier open weights
2026-10-03Verified announcementTRIAL
DwarfStar is an MIT-licensed C inference engine specialized for DeepSeek V4, GLM 5.x and Qwen3.8 on Metal, CUDA and ROCm. It combines project-specific GGUFs, SSD streaming, persistent KV caches, OpenAI- and Anthropic-style APIs and a native coding agent. The project has more than 23,000 GitHub stars but describes itself as fast-changing beta software.
What this means for you
For teams with high-memory workstations, ds4 offers a focused path to run large routed-MoE models locally instead of a generic runtime. Its narrow model support and beta stability require hardware-specific validation before relying on it.
Evidence & related material
Technical details & scores
| Backends | Metal, CUDA and ROCm |
|---|---|
| Models | DeepSeek V4/V4.1, GLM 5.x, Qwen3.8 Flash Next |
| Features | asymmetric quantization, SSD streaming, disk-persistent KV cache, tensor and pipeline parallelism |
| Interfaces | CLI, local APIs and native coding agent |
| License | MIT |
| Adoption | 23,097 GitHub stars and 2,222 forks checked 2026-10-03 |
| Stability | fast-changing beta |
Technical impact 3
Engineering value 4
Adoption signal 5
Maturity 2
Verification cost 4
Risk 3