Redis's creator returns with a new answer
After building Redis, what does Salvatore Sanfilippo (antirez) do next? His answer is DwarfStar (ds4) — a deliberately narrow, self-contained local inference engine. As the README puts it:
> DwarfStar is self-contained and deliberately narrow, not a general GGUF runner.
The Redis philosophy, continued
Redis succeeded by doing one thing to the extreme — being the fastest in-memory cache, not a do-everything database. Twenty years later, antirez applies the same philosophy to inference engines. ds4 supports only three models: DeepSeek V4 Flash, GLM 5.2, and DeepSeek V4 PRO — but runs them extremely well. When better open weights appear, the model list can be swapped.
Why not just use vLLM or llama.cpp?
General GGUF runners make many compromises to support all models. Meanwhile, vLLM has dropped support for older GPUs (Ada Lovelace architecture), turning many enterprise L40S cards into shelfware. ds4's opportunistic strategy: follow the best open weights, support few models, but run them fastest.
Technical highlights
Backends
- Metal: primary target for 96GB+ Macs; SSD streaming for smaller machines
- NVIDIA CUDA: including multi-GPU systems and DGX Spark
- ROCm: Strix Halo systems (e.g., Framework Desktop)
- 120 t/s aggregate generation speed
- 2000 t/s prefill
- SSD streaming: run on a MacBook even when RAM is insufficient
- Tensor parallelism: two MacBook M5 Max / M3 Ultra machines running 4-bit DeepSeek Flash or GLM 5.2 via RDMA
- Pipeline parallelism: chain systems to pool RAM for larger models
- Micro batching: decode and generation batched separately
- colibrì (744B parameters in ~1300 lines of C): small, precise C code still has a place in AI infrastructure
- Euclid-MCP (reasoning offloading): specialized tools for specialized tasks
- vLLM dropping old GPUs: general-purpose compromises are specialized systems' opportunities
- General GGUF runner = Swiss Army knife: does everything, excels at nothing
- ds4 = surgical scalpel: one cut, made perfectly
- Redis then = fastest in-memory cache, no OLTP
- ds4 now = fastest DeepSeek Flash, no other models
- Project: https://github.com/antirez/ds4
- Author: Salvatore Sanfilippo (antirez, creator of Redis)
- License: MIT
- Backends: Metal / CUDA / ROCm
Measured performance
Multi-user session test on 8x L40S NVIDIA cards:
This turns "old cards abandoned by vLLM" into a multi-user LLM server — possibly ds4's most practical selling point for enterprises.
Other features
Self-contained
Model loading, prompt rendering, tool calling, KV state, HTTP server, coding agent — all built and tested together as one integrated piece of engineering.
A candid disclosure of AI-assisted development
The README contains a rare, honest statement:
> This software is developed with strong assistance from GPT 5.5, 5.6, Claude Fable and with humans leading the ideas, testing, and debugging. We say this openly because it shaped how the project was built. If you are not happy with AI-developed code, this software is not for you.
Humans lead ideas, testing, and debugging; AI drives code generation — a new engineering paradigm of human architect + AI programmer.
Crediting llama.cpp
ds4 does not link GGML, but the README states:
> This project would not exist without llama.cpp and GGML.
Some source-level fragments (GGUF quantization layout, CPU quant/dot logic, specific kernels) are retained or adapted under MIT, with GGML authors' copyright notices preserved in the LICENSE file — a respectful independent implementation, not a fork.
Why deliberate narrowness wins
ds4 resonates with recent work:
The common theme: in the era of "bigger is better," deliberate narrowness is an underrated strategy.
Analogy
Conclusion
antirez is swimming against the mainstream. In an arms race of "more models, more parameters," ds4's answer is "fewer models, faster execution." Redis proved deliberate narrowness wins in databases; ds4 is testing whether it holds for inference engines. If it succeeds, it offers a counterintuitive template for AI infrastructure: in an age of generalization, being deliberately narrow may be a competitive advantage.
---