avx2

kimi-k3-in-c

+71

A 2.78-trillion-parameter Kimi K3 running inference on a single CPU in 8.24 GB of RAM. Portable C99: no BLAS, no framework, no GPU.

simdjson

+29

Parsing gigabytes of JSON per second : used by Facebook/Meta Velox, the Node.js runtime, ClickHouse, WatermelonDB, Apache Doris, Milvus, StarRocks

strata-nvfp4

+18

Strata fork: Qwen3.8-Flash-Next 125B MoE in NVFP4 on one RTX 20-50 card (12 GB+) and 64 GB of RAM or more. Setup installs our GPTQ quants (huihui abliterated, OrcaRouter) from Hugging Face. CPU experts on AVX-512 or AVX2 (+AVX-VNNI), all experts in RAM from 92 GB, a low-RAM mode below, vision, W4A8 prompts. Windows / CUDA 13.

janas

+18

37 tok/s from Qwen3-Next-80B (48 GB) on a 32 GB laptop, 24 without drafts vs 2.8 for llama.cpp; reading 62 vs 16 tok/s. Faster than llama.cpp on all 13 models compared in Q4_K_M. LLMs bigger than your RAM, pure C, CPU or GPU.

google/highway

+10

Performance-portable, length-agnostic SIMD with runtime dispatch

CTranslate2

+10

Fast inference engine for Transformer models

simde

+7

Implementations of SIMD instruction sets for systems which don't natively support them.

simdutf

+5

Unicode routines (UTF8, UTF16, UTF32) and Base64: billions of characters per second using SSE2, AVX2, NEON, AVX-512, RISC-V Vector Extension, LoongArch64, POWER. Part of Node.js, WebKit/Safari, Ladybird, Chromium, Cloudflare Workers, Ghostty and Bun.

toks

+3

exact tokenizers ids, as fast as the hardware allows