Public engine component of Project Den (under active development, not released). llama.cpp-based long-context GGUF inference with MTP speculative decoding, quantized KV, adaptive KV streaming, and MoE expert placement on Blackwell sm_120a. build2 stable; build2.1 next.
cuda moe quantization blackwell kv-cache long-context llama-cpp local-llm speculative-decoding nvfp4 expert-offloading sm120a
-
Updated
Oct 23, 2026 - C++