FareedKhan-dev / kimi-k3-in-c
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.
FareedKhan-dev / kimi-k3-in-c
NX-AI / KronLSTM
Flash-linear-attention compatible KronLSTM model and kernels by NXAI.
fla-org / flash-linear-attention
🚀 Efficient implementations for emerging model architectures
Intellindust-AI-Lab / GTR
GTR: Gated Token Recurrence for Efficient Dense Prediction
BlinkDL / RWKV-LM
RWKV (pronounced RwaKuv) is an RNN with great LLM performance, which can also be directly trained like a GPT transformer (parallelizable). We are at RWKV-7 "Goose". So it's combining the best of RNN and transformer - great performance, linear time, constant space (no kv-cache), fast training, infinite ctx_len, and free sentence embedding.
thu-ml / SLA
SLA: Beyond Sparsity in Diffusion Transformers via Fine-Tunable Sparse–Linear Attention
HaoyiZhu / Piecewise-Taylor-Attention
Piecewise-Taylor Attention
TimRots / kimi3
Independent from-scratch implementation of the Kimi K3 architecture (arXiv:2607.24653v1): KDA, NoPE, latent-space MoE, and the systems co-designs. Table 1 reproduced to 0.09%.
lucidrains / bdh-cq
Implementation of BDH-CQ: In-Context Learning with Recurrent Latent Reasoning, proposed by Pathway Research