DwarfStar 4 (ds4)
dwarfstar.sh- Category
- Developer Tools
- Rank
- No. 2834Tools index
Previous survey · No. 2583 ·
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Models: Train & Run
- Interfaces
- CLI · API
- Date
About
DwarfStar 4 is a C inference engine by antirez for running DeepSeek V4 and V4.1 Flash, GLM 5.x and Qwen3.8 Flash Next locally on Metal, CUDA and ROCm machines. It uses asymmetric 2-bit quantization of the routed experts so large mixture-of-experts models fit on high-memory hardware, and it persists the KV cache to SSD so long prompt prefixes can be reloaded after a restart instead of recomputed. It ships a CLI, an OpenAI and Anthropic compatible API server, and a native coding agent.
At a glance
DwarfStar 4 runs selected large models locally with a C inference engine and hardware-specific backends. It is worth evaluating if you have compatible hardware and want to run a supported model without a hosted inference endpoint.
Good fit
Local inference experimenters with substantial memory and a workload that fits the supported model and quantization options.
Consider something else when
Pass if your hardware falls below the documented requirements or you need broad model compatibility from one engine.
Start here
Check your exact hardware and model against the support table, then compare published prompt and generation speeds with what your workload needs.
Cost and requirements
The beta documents Metal, CUDA and ROCm paths for specific systems. Model weights and storage can be large, and required memory depends on the model and configuration.
Limits to know
Lower-bit weights and disk streaming trade memory against speed or output quality. Check measurements for the same backend and model before extrapolating.
Sources checked 2026-11-05. Source-based assessment.
Tags
Media

Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.