Fast Qwen3.8-27B for one RTX 3090. 200k context, 100+ tok/s. Custom CUDA decode engine, OpenAI-compatible API.
-
Updated
Nov 18, 2026 - Python
Fast Qwen3.8-27B for one RTX 3090. 200k context, 100+ tok/s. Custom CUDA decode engine, OpenAI-compatible API.
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.
CPU inference engine in Rust for Qwen3.8-27B: run the full Q4_K_M model on a 16 GB laptop by streaming what does not fit, instead of dropping to a worse quant. Hand-written AVX2 K-quant kernels, gated DeltaNet + GQA, identical token ids at every memory budget and thread count.
Tilde Research Challenge - DeltaNet
Run 27B-parameter Qwen3.8 locally on a laptop CPU with native C inference, no GPU or Python, at 2.52 token/s.
Distributed Engrammics — gradient-free transfer of a persistent fast-weight associative state between model instances (DeltaNet, RWKV-7). Paper + reproducible code.
Two fast-weight experiments (FENRIR and FREKI) on whether a cheap recurrent layer can chain multi-hop lookups inside one forward pass, without attention. Honest results, builds on prior art.
Delta Coin(DTC) erc20 token
Uno-style diffusion drafting and transactional Gated DeltaNet verification on Qwen3.5/MLX
Nested memory states for RNNs, with learned promotion between levels
Continuous latent reasoning (Coconut) ported to Qwen3.5's hybrid gated-DeltaNet + full-attention architecture. At 500-example LoRA scale on Kaggle T4s, latent tokens gave no measurable benefit over a same-curriculum no-thought control; ablations and raw logs included.
The Delta Rule Lab — error-correcting memory in linear attention (DeltaNet / Gated DeltaNet). Real in-browser autodiff, live MQAR training, gradient-checked.
Zero-Copy C++20 / Objective-C++ / Metal / Accelerate Kernel Toolkit for PyTorch MPS on Apple Silicon
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
Controlled benchmark of memory mechanisms for transformers, built on nanochat
To associate your repository with the deltanet topic, visit your repo's landing page and select "manage topics."