CORE 01
Asymmetric 2-bit quantization
Compress the routed experts, keep critical shared paths precise. That is how the supported routed-MoE builds fit their target machines.
DS4 · LOCAL FRONTIER INFERENCE
Run frontier open weights locally with ds4.
DwarfStar 4 is a narrow C inference engine for high-memory Mac, CUDA and ROCm machines. It supports DeepSeek V4 and V4.1 Flash, GLM 5.x and Qwen3.8 Flash Next, with text and vision models, local APIs, a CLI and a native agent in one stack.
SUPPORTED: DEEPSEEK V4 / V4.1 + GLM 5.x + QWEN3.8 · MIT LICENSE · C / METAL / CUDA / ROCM · QWEN ON 64GB
PRINCIPLE · LOCAL MODEL STACK
PHASE 1 · THE GIANT
A 284-billion-parameter star
DeepSeek V4 Flash is a large mixture-of-experts model. The usual path is remote serving; ds4 starts from the opposite constraint.
PHASE 2 · THE COLLAPSE
Compressed, not lobotomized
Asymmetric quantization targets the routed experts while preserving critical paths. The model becomes practical on high-memory machines.
PHASE 3 · THE DWARF STAR
Dense, resident, yours
The local engine exposes a CLI, HTTP APIs and a native agent, all sharing the same model state and cache.
SCROLL ▾
DESIGN CHOICES
Local frontier inference, narrow on purpose
Not a generic GGUF runner. ds4 follows a small, opportunistic set of model families and validates each supported layout end to end.
CORE 02
KV cache as a disk citizen
Save long prefixes to SSD and resume by prompt hash. Restarts do not have to mean full re-prefill.
CORE 03
One engine, three interfaces
Use ./ds4 for chat, ./ds4-server for local
APIs and ./ds4-agent for persistent coding sessions.
- SSD STREAMING
- TENSOR PARALLELISM
- SESSION BATCHING
- DSPARK + MTP
- VISION INPUT
ARCHITECTURE
How the ds4 stack fits together
Project GGUFs, a self-contained engine and agent-facing interfaces, checked against official model outputs.
RUNTIME MAP · SIMPLIFIED. SEE ARCHITECTURE NOTES FOR THE FULL DRAWING.
RUN IT
Run ds4 in three steps
Download the project GGUF, build for your backend, then start the CLI or server. Generic GGUF files are not the target.
STEP 1 · FETCH THE WEIGHTS
$ git clone https://github.com/antirez/ds4
$ cd ds4 && ./download_model.sh ds4f-q2
STEP 2 · BUILD FOR YOUR BACKEND
$ make
$ make cuda-spark
STEP 3 · TALK TO IT
$ ./ds4
$ ./ds4-server --ctx 100000
FIT CHECK
ds4 hardware fit: local, streamed and distributed
Pick your platform and memory: get a conservative starting path and understand which execution modes apply.
✓ Runs well
V4 Flash Q2 is the baseline. At 128 GB, GLM 5.3 Q2 and Qwen Q4 also fit; V4.1 Q2 streams from SSD.
./download_model.sh ds4f-q2 && make
REF · M5 MAX 128GB · 32K CTX: 34.4 T/S GEN · 557 T/S PREFILL
Estimates from the ds4 benchmark table. Full guide in Hardware and Installation.
BENCHMARKS
ds4 benchmarks: prefill and generation
Reference rows from upstream. Read prefill and generation separately, especially for long-context agent workloads.
| Machine | Context | Prefill t/s | Generation t/s |
|---|---|---|---|
| M5 Max, 128 GB | q2 · 2,048 tok | 790.2 | 39.4 |
| M5 Max, 128 GB | q2 · 65,536 tok | 398.5 | 27.6 |
| DGX Spark, 128 GB | q2 · 2,048 tok | 825.8 | 18.1 |
| DGX Spark, 128 GB | q2 · 65,536 tok | 823.0 | 13.8 |
API & AGENTS
Use ds4 from Codex, Claude Code and OpenCode
ds4-server speaks OpenAI and Anthropic-style APIs, so local coding agents can connect to your own machine with a base URL.
Own your local AI inference.
Start with the quickstart, check the hardware matrix, then connect your editor, agent or API client to the local server.