面向手机端的高性能神经网络推理框架，针对 ARM 平台深度优化，让 Caffe、ONNX 等模型能在 Android/iOS 上低延迟运行。
基于 TurboQuant 量化算法、用 Rust 实现并带 Python 绑定的向量索引，面向 RAG 和嵌入检索场景，提供高压缩比、低内存的近似最近邻搜索。
用纯 C99 从零实现的 Kimi K3 推理引擎，单 CPU、8.24 GB 内存即可跑 2.78 万亿参数模型，不依赖 BLAS、框架或 GPU。
…cQuantizedInnerProduct32fRef.</li <li C++ wrapper Simd::SynetDynamicQuantizedInnerProduct32f.</li <li Base implementation, SSE4.1, AVX2, AVX-512BW, AVX-512VNNI, AMX-INT8,
…thms</h4 <h5 New features</h5 <ul <li C++ wrapper Simd::SynetScale16b.</li <li C++ wrapper Simd::CopyFrame.</li <li Base implementation, SSE4.1, AVX2, AVX-512BW, AVX-512V
…ions of class ImageBmpLoader.</li <li C++ wrapper Simd::SynetAdd16b.</li <li C++ wrapper Simd::SynetQuantizedAdd.</li <li C++ wrapper Simd::SynetQuantizedMul.</li <li C++
…include by @yonik in https://github.com/fast-pack/SIMDCompressionAndIntersection/pull/25 fix buggy use of 32 bit ctzl that resulted in crashes with recent MSVC by @yonik
…per-segment query scale. Deferred to v0.8.1. WASM SIMD port — requires first un-stubbing turbo kv attention in quant.h single-header. Tests 35/35 passing. New TurboKVRegr
…core allocations to mimalloc for lower overhead - simdjson log parsing — Faster deeplog deserialization via simdjson - LZ4 compressed deeplog logs — Reduced on-disk deepl