
DwarfStar 4
Runs large models such as DeepSeek V4 Flash and GLM 5.3 Flash on Apple Silicon, NVIDIA and AMD.
What it does better
DeepSeek Flash at 2-bit, about 81 GiB, decodes at 34.36 tok/s with a 32,768-token context on a 128 GB Mac
GLM 5.3 Flash at Q4_K, 177.77 GiB and larger than the 128 GB of memory, runs from the SSD at 11.9–14.9 tok/s (three-run median, 2K prompt plus 1K append, 128 generated tokens)
Trade-off: the expert weights are read from the SSD, so it needs a fast local SSD, and the speed depends on the prompt.
Two 128 GB Macs linked over RDMA run DeepSeek V4 Flash Q4 or MXFP4, or GLM 5.3 Flash Q4, with tensor parallelism that splits the experts between them
Trade-off: the link between the machines has no authentication or encryption, so it needs a trusted network.
DeepSeek V4 Flash runs on eight NVIDIA L40S cards (Ada Lovelace) at about 126 tok/s aggregate with 16 sessions; the author says other implementations of these models need newer GPUs
Trade-off: the 126 tok/s is total throughput across sessions, not the speed of one user.
Benchmarks
The author hasn't published benchmarks yet.
Listing
Listed: Passed all seven checks.
Security review
Not reviewed yet.
Install
To install it, follow the project's own instructions on GitHub. The engine runs with your user's permissions. macOS doesn't sandbox it, so install only engines you trust.
Something wrong or out of date? Write to help@kalsa.io, or ask us to remove this page.