DwarfStar brings DeepSeek V4 local inference

DwarfStar makes a compelling bet that model-specific runtimes can outperform generic inference stacks for local agent workloads, though its narrow model support and high memory requirements limit its audience.

• –Native KV-cache handling and on-disk session persistence target long-running coding-agent sessions
• –Metal support makes high-end Apple Silicon a practical target for frontier-scale local models
• –CUDA and ROCm support broaden the project beyond Mac hardware, including DGX Spark and AMD systems
• –Its deliberate specialization enables tighter validation and optimization than general GGUF runners
• –Beta quality, limited model compatibility, and single-session server inference remain meaningful production constraints