Redis Creator Releases ds4: C-Based Engine for Local Frontier LLM Inference

The open-source tool ds4, developed by the creator of Redis, enables local execution of frontier open-weight models including DeepSeek V4, V4.1 Flash, GLM 5.x, and Qwen3.8 Flash Next. The engine is written in C and supports Apple Silicon (Metal), NVIDIA (DGX Spark/CUDA), and AMD (Strix Halo/ROCm) platforms with memory configurations ranging from 32 GB to 512 GB. Key features include asymmetric 2-bit quantization using imatrix, KV cache management that survives restarts by streaming to SSD, and unified interfaces for CLI, OpenAI/Anthropic-compatible APIs, and coding agents. Benchmarks on an M5 Max (128 GB) show 34.4 tokens/s generation and 557 tokens/s prefill at 32K context, while DGX Spark (128 GB) achieves 18.1 tokens/s generation and 825.8 tokens/s prefill at 2K context. The project integrates with tools like OpenCode, Claude Code, and Codex CLI.