int4

SOTA low-bit LLM quantization (INT8/FP8/MXFP8/INT4/MXFP4/NVFP4) & sparsity; leading model compression techniques on PyTorch, TensorFlow, and ONNX Runtime

A simple and effective post training quantization toolkit for high-accuracy low-bit LLM inference|简洁且高效的后训练量化工具包

row-major matmul optimization

An innovative library for efficient LLM inference via low-bit quantization

Pure-Rust, CPU-only OCR engine for Baidu Unlimited-OCR (a DeepSeek-OCR-derived 3B MoE VLM). Five-model zoo, custom int8 kernels, no ML framework, no Python, no GPU.

[ICML 2026] GRACE: deployable INT4 VLM via quantization-aware distillation.

Hand-written NVFP4 W4A16 CUDA kernels for Volta

Qwen3.8-Flash-Next on 2× RTX 3090: with 128 GB RAM up to 4,191 tok/s prefill · 111.5 tok/s decode (131K prompt), full 256K window at 2,865 tok/s; with 64 GB RAM 3,410 tok/s prefill · 84 tok/s decode. New: opt-in uncensored mode (runtime abliteration, no new weights).

600 KB WASM runtime for Cactus Compute's Needle AI tool-calling models, Needle 3, 2 and 1 from one build. Browser, Cloudflare Workers, Node.js, Python, C FFI and no_std. Token-exact with the JAX reference. No backend, no API key.

PrismQuant: Optimal Null-Space Rotations for Grouped Quantizers

GLM-5.2, a 744 billion parameter mixture of experts model, in a pure C inference engine: quantized to int4, experts streamed from disk, deployed and benchmarked. Generates in 16 GB of RAM.

Serving 4-bit Qwen3.8-27B on a single DGX Spark (GB10): 75 tok/s single-stream, 246 tok/s aggregate at 8-way concurrency. NVFP4 vs MixedInt4-AutoRound vs the FP8 baseline, measured on one harness — including why the quantization advantage collapses to +0.2% by c16.

Optimized local serving engine for Kimi-Linear-48B: INT4 quantizer, fused decode kernels for a measured 3.18x, and an OpenAI-compatible server. Ships with k3, a bridge that detects the client per request so Claude Code, Codex, Cline, Aider and opencode all work unchanged.

🧬🔍 Vecgo is a pure Go, embeddable, hybrid vector database designed for high-performance production workloads. It combines commit-oriented durability with HNSW + DiskANN indexing for best-in-class performance.

⚡️ The fastest way to run local LLMs on Apple Silicon — sub-second model loads, beats Ollama on throughput, tail latency, and full-response time. OpenAI/Ollama-compatible. No cloud, no API keys.

Run Qwen3.6-27B on four Tesla V100s at 366 tok/s using hand-written NVFP4 CUDA kernels and chain-MTP speculation.

Single-stream LLM inference engine for Qwen3.8-27B on one RTX 5090 — int4 quantization + speculative decoding + custom Triton/CUDA kernels; 229–379 tok/s vs llama.cpp/SGLang/vLLM at batch size 1.

Research and training stack for AVA — a tool-using, memory-aware virtual assistant targeting 4 GB VRAM. Spans custom transformers, verifier-RL, external memory, multi-domain benchmarks, and Gemma 4 inference optimization.

Three ComfyUI nodes that show the real runtime precision of MODEL, CLIP, and VAE objects.

Run Qwen3.8-Flash-Next on ONE RTX 3090 (24 GB) + 64 GB RAM: 128K context, up to 2,100 tok/s prefill, 43–51 tok/s decode. vLLM runtime with hot MoE experts on the GPU and cold experts computed on the CPU, INT8 KV cache, Docker, OpenAI-compatible API.