turboquant

A vector index built on TurboQuant, written in Rust with Python bindings

Self-hosted AI agent OS. Your memory, chat, agents, and files stay on hardware you own, offline by default, cloud by choice. Offline AI memory (taOSmd), self-hosted multi-framework group chat, a full web desktop + app store, and auto-clustering across the consumer hardware you already have (Orange/Raspberry Pi, Mac mini, gaming PC).

llama.cpp fork with TurboQuant WHT-rotated KV cache & weight compression + Gemma 4 MTP and Qwen 3.6 NextN speculative decoding (+30-50% throughput).

LLM inference with 7x longer context. Pure C, zero dependencies. Lossless KV cache compression + single-header library.

Qwen 3.8 27B ROCmFP4 on AMD Strix Halo (Ryzen AI Max+ 395). Up to 36 tok/s via MTP Speculation, TurboQuant & Mesa RADV Wave64.

llama.cpp fork for significantly improved performance on Ampere (especially RTX 3090 / 3090 Ti): example: 95+ tok/s over a 100K-token generation at temperature 1 for qwen3.8, supporting full 262K ctx. Supports low memory (12gb) configs for 200k+ ctx

Based on the implementation of Google's TurboQuant (ICLR 2026) — Quansloth brings elite KV cache compression to local LLM inference. Quansloth is a fully private, air-gapped AI server that runs massive context models natively on consumer hardware with ease

TurboQuant KV cache compression for MLX with fused Metal kernels. 4.6x compression at 98% FP16 speed.

Qwen4-Exp (Qwen3.8-Flash-Next) SWA + MTP + PLE-on-iswa — bounded deep-context decode w/ long-range recall, TBQ4 4.25 bpv KV, DSpark + RotorQuant. Main (master) = Qwen4-Exp build; the Qwen35 SWA hybrid lives on the qwen35 branch. RTX 4090.

Extreme weight + KV cache compression for LLMs on Apple Silicon (MLX implementation of Google's TurboQuant)

Native Windows vLLM: 0.27.1 Latest; 0.29.0 Python 3.14/CUDA 13.2 release (CPU TorchAudio), plus Python 3.13/CUDA 13.0 prerelease. PyTorch 2.13, FlashAttention/Rust, OpenAI serving, 10 KV formats, Multi-TurboQuant, experimental CPU/NVMe KV offload. No WSL or Docker.

llama.cpp fork with TurboQuant quantization (turbo2/3/4) and TriAttention GPU-accelerated KV cache pruning. 75 tok/s on Qwen3-8B / RTX 3080.

TurboQuant 3-bit KV-cache quantization for llama.cpp

TurboQuant KV cache compression plugin for vLLM — asymmetric K/V, 8 models validated, consumer GPUs

TurboQuant-compatible vector search plus graph memory for constrained RAG.

First open-source TurboQuant KV cache compression for LLM inference. Drop-in for HuggingFace. pip install turboquant.

llama.cpp fork optimized for NVIDIA DGX Spark / GB10 (Blackwell, SM 12.1) — TurboQuant weights + KV, NVFP4, DFlash MTP

Minimal, zero-dependency LLM inference in pure C11. CPU-first with NEON/AVX2 SIMD. Flash MoE (pread + LRU expert cache). TurboQuant 3-bit KV compression (8.9x less memory per session). 20+ GGUF quant formats. Compiles to WASM.

Unified KV-cache compression for LLM inference: 12 Python-native methods, guarded add-on composition and routing, analytical capacity simulation, Godzilla KVarN/TriAttention with SM86/SM89 qualification, CUDA weight sharing, and multi-GPU planning.

DGX Spark research and tests - containers, benchmarks, and investigation notes for running models on GB10 (SM 12.1)