antirez

ds4 DeepSeekV4 Flash推理引擎

Mac本地推理引擎

22.2k 2.1k 172 632 388

核心框架 AI平台模型推理引擎 macOS

(0 评价)

ds4 DeepSeekV4 Flash推理引擎

GitHub Trend · Star / Watch / Fork

使用 ds4，你可以：

由 Redis之父antirez 开发的专用高性能推理引擎，仅为 DeepSeek V4 Flash 模型设计，针对 Apple Metal 架构做极致底层优化。采用模型定制化内核与内存管理，支持 Q2/Q4 量化，可在 128GB 内存 Mac 上高效运行 284B 参数 MoE 模型，显著降低通用框架开销，兼顾吞吐与延迟。

用户评论 (0)

Github榜单

日榜

2026-08-03

rank25

267

ds4 DeepSeekV4 Flash推理引擎

ds4 DeepSeekV4 Flash推理引擎

用户评论 (0)

#41.Critical Security Vulnerabilities: Buffer Overflow and Shell Command Injection

#491.Add optional expert bundle sidecar for contiguous MoE expert reads

#114.q4-imatrix weights fail to load: "expected IQ2_XXS expert tensors"

#78.ds4: CUDA q8 fp16 cache alloc failed (xx.xx MiB): out of memory

#126.speed benchmark on M3 Max Macbook Pro with different fan control.

#34.CUDA work in progress: info

#46.FYI: Works with 96 GB as well

#127.tool call status messages differ vs. llama.cpp

#62.KV prefix cache may miss final response prefix in pi tool sessions

#475.Strix Halo - 128GB with auto or forced 96GB for GPU fails

项目发展历程

Github榜单

推理引擎排行

产品推荐

欢迎回来

ds4 DeepSeekV4 Flash推理引擎

ds4 DeepSeekV4 Flash推理引擎

为ds4打分

收藏到合集

用户评论 (0)

#41.Critical Security Vulnerabilities: Buffer Overflow and Shell Command Injection

#491.Add optional expert bundle sidecar for contiguous MoE expert reads

#114.q4-imatrix weights fail to load: "expected IQ2_XXS expert tensors"

#78.ds4: CUDA q8 fp16 cache alloc failed (xx.xx MiB): out of memory

#126.speed benchmark on M3 Max Macbook Pro with different fan control.

#34.CUDA work in progress: info

#46.FYI: Works with 96 GB as well

#127.tool call status messages differ vs. llama.cpp

#62.KV prefix cache may miss final response prefix in pi tool sessions

#475.Strix Halo - 128GB with auto or forced 96GB for GPU fails

项目发展历程

Github榜单

推理引擎排行

产品推荐