deltanet

Apple Neural Engine (ANE) LLM inference engine — reverse-engineered private APIs, Metal GPU shaders, hybrid ANE+GPU+CPU on Apple Silicon. 32 tok/s matching llama.cpp, 3.6 TFLOPS fused ANE mega-kernels.

Expanding linear RNN state-transition matrix eigenvalues to include negatives improves state-tracking tasks and language modeling without added training or inference costs.

DeltaProduct is a new linear recurrent neural network architecture that uses products of generalized Householder matrices as state-transition matrices enabling better expressivity, state-tracking, and language modeling performance than DeltaNet.

Rust-native MoE inference runtime with custom CUDA kernels for Blackwell GPUs. Includes DFlash speculative decoding, multi-tier Engram memory, and entropy-adaptive routing. Targets Qwen3.5-35B-A3B on a single RTX 5060 Ti 16GB.

Zero-framework, 100% CPU inference engine in C11 (AVX2 + FMA) for hybrid Qwen 3.8-27B (Gated DeltaNet + GQA Attention) with native speculative decoding.

Tilde Research Challenge - DeltaNet

Distributed Engrammics — gradient-free transfer of a persistent fast-weight associative state between model instances (DeltaNet, RWKV-7). Paper + reproducible code.

Nested memory states for RNNs, with learned promotion between levels

The Delta Rule Lab — error-correcting memory in linear attention (DeltaNet / Gated DeltaNet). Real in-browser autodiff, live MQAR training, gradient-checked.

Uno-style diffusion drafting and transactional Gated DeltaNet verification on Qwen3.5/MLX

minGRU (Feng et al.) in PyTorch, plus a measured ladder of state-tracking recoveries — signed, rotation, Givens, and delta (DeltaNet-style Householder products) mixers with time decay, all under one parallel scan — with fused Triton GPU kernels as an optional backend

Delta Coin(DTC) erc20 token

Controlled benchmark of memory mechanisms for transformers, built on nanochat

Run 27B-parameter Qwen3.8 locally on a laptop CPU with native C inference, no GPU or Python, at 2.52 token/s.