Redis creator's ds4 runs DeepSeek V4 locally behind a local API
An MIT C engine, 2-bit quantized, serves an OpenAI/Anthropic-style local API — plus Opus 5.5 oil-paints in code and unearths a lost dodo record.
Copy markdown
A local inference engine from the Redis creator
antirez's ds4 (DwarfStar 4) is an MIT-licensed C engine that runs DeepSeek V4 and V4.1 Flash, GLM 5.x, and Qwen 3.8 Flash Next — text and vision — on a 64GB+ Apple Silicon Mac, a CUDA box or DGX Spark, or an AMD Strix Halo desktop. Asymmetric 2-bit quantization keeps the weights in memory; an M5 Max clocks ~790 tok/s prefill and ~39 tok/s generation. The ds4-server exposes OpenAI- and Anthropic-compatible endpoints, so you can point existing agent code at a local model and zero out that API bill, and the KV cache persists to SSD so sessions resume instead of re-prefilling.
Opus 5.5 paints oil in code, no image generator
Stillwet hands Claude Opus 5.5 a simulated oil-paint canvas it drives by writing code — bristle brushes, wet paint, drying, and glazing on linen — and lets it work iteratively at a "virtual easel," stepping back to inspect a passage before adding more. The strokes are programs, not diffusion output, and the source is on GitHub: a working template if you want to build code-driven creative tools instead of prompt-to-image.
Opus 5.5 digs a lost dodo sighting out of the archives
A historian used Opus 5.5 to surface a previously-unnoticed 17th-century eyewitness account of the dodo, buried in archival text. It's a concrete case for long-context models as research assistants over primary sources — the unlock for anyone building on document corpora is reasoning across raw archives, not just answering from a chat window.