@axiaisacat: Redis creator antirez drops another hardcore project: ds4. Not just another GGUF runner, but a local inference engine specifically written for DeepSeek V4 Flash: Metal / CUDA 2-bit quantization 1M context KV ...

X AI KOLs Timeline Tools
redis antirez local-inference deepseek quantization metal cuda

Summary

Redis creator antirez released ds4, a local inference engine optimized for DeepSeek V4 Flash with 2-bit quantization and support for 1M context KV cache on Metal and CUDA.

Redis creator antirez drops another hardcore project: ds4. Not just another GGUF runner, but a local inference engine specifically written for DeepSeek V4 Flash: Metal / CUDA 2-bit quantization 1M context KV cache can be saved to disk Targeting high-end Macs and DGX Spark The key point is not just 'can run', but to make a local large model into a complete and usable engineering closed loop. Local AI https://t.co/c2TFtHfQwX
Original Article
View Cached Full Text

Cached at: 05/14/26, 04:40 PM

Redis author antirez has dropped another hardcore project: ds4.

Not just another GGUF runner, but a local inference engine specifically written for DeepSeek V4 Flash:

  • Metal / CUDA
  • 2-bit quantization
  • 1M context
  • KV cache can be offloaded to disk
  • Targeting high-end Mac and DGX Spark

The focus is not just on “being able to run”, but on turning a local large model into a complete, usable engineering closed loop.

Local AI https://t.co/c2TFtHfQwX

Similar Articles

From the creator of Redis; run LLM locally with ds4

Hacker News Top

Redis creator antirez released ds4 (DwarfStar 4), a narrow C-based local inference engine for high-memory Mac, CUDA and ROCm machines, supporting DeepSeek V4/V4.1, GLM 5.x and Qwen3.8 Flash Next with asymmetric 2-bit quantization, SSD-backed KV cache, and CLI/server/agent interfaces under an MIT license.

antirez/deepseek-v4-gguf

Hugging Face Models Trending

Antirez released GGUF quantizations of DeepSeek V4 Flash specifically tailored for the DS4 inference engine, providing optimized configurations for different RAM sizes and enabling local execution of the large MoE model.