What does the creator of Redis do next? Salvatore Sanfilippo (antirez)'s answer is DwarfStar (ds4) — a deliberately narrow, local inference engine. The project has been climbing GitHub Trending (~385 stars/day), and it represents antirez's new answer twenty years after Redis.
Continuing the Redis Philosophy
Redis succeeded by doing one thing extremely well — being the fastest in-memory cache rather than a do-everything database. antirez applies the same philosophy to inference:
> DwarfStar is self-contained and deliberately narrow, not a general GGUF runner.
It is not a general GGUF runner. It supports only three models: DeepSeek V4 Flash, GLM 5.2, and DeepSeek V4 PRO — but runs them as fast and as well as possible.
Why Not Just Use vLLM or llama.cpp?
The local inference engine space is crowded. antirez's argument: general-purpose GGUF runners make too many compromises to support every model. vLLM no longer supports older GPUs (Ada Lovelace architecture), which turns many enterprises' L40S cards into dead weight. ds4's opportunistic strategy: follow the best open weights, run only a few models, and run them fastest. Models can be swapped as better open weights appear.
Technical Highlights
Backends
- Metal: primary target for Macs with 96GB+ RAM; 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 throughput
- 2000 t/s prefill
- SSD streaming: run on a MacBook even with insufficient RAM
- Tensor parallelism: two MacBook M5 Max / M3 Ultra machines running 4-bit DeepSeek Flash or GLM 5.2 over RDMA
- Pipeline parallelism: chain multiple systems to pool RAM for larger models
- Micro batching: separate batching for decode and generation
- colibrì (744B parameters on 1300 lines of C): small, precise C code still has a place in AI infrastructure
- Euclid-MCP (outsourced reasoning): let specialized tools do specialized jobs instead of one universal system doing everything
- vLLM dropping old GPUs: general systems' compromises are specialized systems' opportunities
- General GGUF runners = Swiss Army knife: does everything, excels at nothing
- ds4 = a scalpel: makes one kind of cut, but cuts most precisely
- Redis then = fastest in-memory cache, no OLTP
- ds4 now = fastest DeepSeek Flash, no other models
Measured Performance
Multi-user session tests on 8x L40S NVIDIA cards:
This turns "older cards abandoned by vLLM" into a multi-user LLM server — possibly ds4's most practical selling point for enterprises: no need to throw away existing GPUs.
Other Features
Self-Contained
Model loading, prompt rendering, tool calling, KV state, HTTP server, and a coding agent — all built and tested together. Not assembled from components, but one integrated engineering effort.
A Candid Statement on AI-Assisted Development
The ds4 README contains an unusually candid disclosure:
> 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.
antirez doesn't hide the AI assistance — he openly states it shaped the project. Humans lead ideas, testing, and debugging; AI leads code generation. A new engineering paradigm: human architect + AI programmer.
Credits to llama.cpp
ds4 doesn't link GGML, but the README states:
> This project would not exist without llama.cpp and GGML.
Some source-level snippets (GGUF quantization layout, CPU quant/dot logic, specific kernels) are retained or adapted under MIT. antirez even keeps the GGML author's copyright notice in the LICENSE file. This is open-source community interaction at its best: standing on predecessors' shoulders, openly crediting, and preserving copyright. Not a fork — an homage-style independent implementation.
Engineering Insight: The Win of Deliberate Narrowness
ds4's design philosophy resonates with recent work:
Together these point to one theme: in the "bigger and stronger" AI narrative, "deliberately narrow" is an underrated strategy.
Analogy
Closing
antirez chose a path opposite to the mainstream. In the arms race of "support more models, run bigger parameters," ds4 answers with "support fewer models, run faster." Redis proved deliberate narrowness is a winning strategy in databases; ds4 is testing whether it holds for inference engines. If it succeeds, it's not just another inference engine — it's a counterintuitive template for AI infrastructure: in an era of generalization, deliberate narrowness may be an underrated competitive edge.
---
Project: https://github.com/antirez/ds4 Author: Salvatore Sanfilippo (antirez, creator of Redis) License: MIT Backends: Metal / CUDA / ROCm