NVIDIA has released the Nemotron-TwoTower-30B-A3B-Base-BF16 model, which is built upon the Nemotron 3 Nano 30B-A3B backbone. This architecture diverges from standard autoregressive models by utilizing a frozen context tower alongside a diffusion denoiser tower. The system iteratively fills blocks of tokens in parallel rather than generating them strictly one at a time. According to NVIDIA, this default mask-diffusion setup retains 98.7% of the aggregate benchmark quality found in the autoregressive baseline. Despite maintaining high quality, the model achieves 2.42 times its wall-clock generation throughput. The release highlights a novel approach to language modeling that combines diffusion techniques with large-scale language capabilities.
NVIDIA Releases Nemotron-TwoTower-30B-A3B, a Diffusion-Based Language Model
llama.cpp adds support for NVIDIA Nemotron-3-Puzzle-75B-A9B
llama.cpp introduces support for the NVIDIA Nemotron-3-Puzzle-75B-A9B model, enabling inference for this Mixture of Experts architecture. The update implements infrastructure for variable-per-layer expert feed-forward sizes and top-k routing, which are required because the model has 40 MoE layers with distinct n_ff_exp and top-k values.
LordNeel releases GGUF quants for Tencent Hy3 and NVIDIA Nemotron-Labs-Audex
LordNeel has published GGUF quantization sets for Tencent's 295B MoE Hy3 model and NVIDIA's audio-capable 30B hybrid MoE Nemotron-Labs-Audex-30B-A3B, providing full benchmark data and reproducible metrics.
NVIDIA releases Puzzle-75B-A9B, a compressed Nemotron model with 2x throughput
NVIDIA has released the Nemotron-Labs-3-Puzzle-75B-A9B model, a deployment-optimized large language model derived from Nemotron-3-Super-120B-A12B. It utilizes the Iterative Puzzle post-training compression framework to significantly improve inference efficiency for reasoning-heavy and long-context workloads while maintaining strong downstream accuracy.
NVIDIA releases Qwen3.6-27B-NVFP4 model
NVIDIA has released the Qwen3.6-27B-NVFP4 model on Hugging Face.
GLM-5.2 on 4x DGX Spark: Reconstructing Missing Build Steps for MTP Speculative Decode
The author successfully deployed GLM-5.2 with MTP speculative decode on a cluster of four NVIDIA GB10 (DGX Spark) nodes, achieving approximately 9.4 tokens per second. This setup utilizes vLLM with tensor parallelism, ported sparse-MLA Triton kernels, and a deterministic 15% expert pruning to fit AWQ-INT4 weights. A critical finding is that the original Docker image build instructions are incomplete, requiring reconstruction of missing patches for deep_gemm.py and sparse_attn_indexer.py. The author also identified that using any vLLM version other than the specific pinned commit causes real AWQ weights to crash during loading due to CUDA errors. To replicate the environment, users must apply a custom script that bakes in kernels and routes functions to sm12x fallbacks. Performance benefits include roughly double the speed of previous llama.cpp implementations, though inter-node bandwidth remains a bottleneck for dual-rail scaling.