#Mamba2
seriously the nemotron team has to be a bunch of masochists to want to use mamba2 again and again for every single model instead of, i don't know, switching to a better linear/recurrent attention implementation that isn't nearly as much of a nightmare to train and to run? fucking hell
November 12, 2026 at 6:23 PM
🚀 New benchmarks show NVIDIA CUDA’s cuTile primitives slashing exp2 latency in Mamba‑2, boosting Tensor‑Core powered sequence modeling. Jane Street’s tricks make attention faster than ever. Dive in to see the numbers! #NVIDIACUDA #Mamba2 #cuTile

🔗 aidailypost.com/news/nvidia-...
October 20, 2026 at 2:09 AM
Бесплатная модель Nemotron Ultra от NVIDIA: где и как применить

Под Nemotron Ultra сейчас имеется в виду NVIDIA-Nemotron-3-Ultra-550B-A55B: 550 млрд параме…

Подробнее: https://gruzdevv.ru/stati/nemotron-ultra-nvidia-besplatnyj-dostup/

#NVIDIA #NemotronUltra #OpenSourceLLM #LLM #Mamba2 #IT #SaaS
September 26, 2026 at 6:00 PM
Zyphra released Zamba2-VL, an open-source vision-language model family that uses a Mamba2-transformer hybrid architecture to target lower-latency multimodal inference for documents, OCR, counting and edge AI tasks.

The release covers three model sizes: 1.2B, 2.7B and 7B parameters.

#AI
Zyphra’s Zamba2-VL Tests Hybrid AI For Faster Vision-Language Models
Zyphra released Zamba2-VL, an open-source vision-language model family that uses a Mamba2-transformer hybrid architecture to target lower-latency multimodal inference for documents, OCR, counting and edge AI tasks.
stechtimes.com
September 9, 2026 at 3:00 PM
NVIDIA AI Brings Nemotron-3-Nano-30B to NVFP4 with Quantization Aware Distillation (QAD) for Efficient Reasoning Inference

NVIDIA has released Nemotron-Nano-3-30B-A3B-NVFP4, a production checkpoint that runs a 30B parameter reasoning model in 4 bit NVFP4 format while keeping accuracy close to its…
NVIDIA AI Brings Nemotron-3-Nano-30B to NVFP4 with Quantization Aware Distillation (QAD) for Efficient Reasoning Inference
NVIDIA has released Nemotron-Nano-3-30B-A3B-NVFP4, a production checkpoint that runs a 30B parameter reasoning model in 4 bit NVFP4 format while keeping accuracy close to its BF16 baseline. The model combines a hybrid Mamba2 Transformer Mixture of Experts architecture with a Quantization Aware Distillation (QAD) recipe designed specifically for NVFP4 deployment. Overall, it is an ultra-efficient NVFP4 precision version of Nemotron-3-Nano that delivers up to 4x higher throughput on Blackwell B200.
nexttech-news.com
February 2, 2026 at 7:47 AM
TII Abu-Dhabi Released Falcon H1R-7B: A New Reasoning Model Outperforming Others in Math and Coding with only 7B Params with 256k Context Window

Technology Innovation Institute (TII), Abu Dhabi, has released Falcon-H1R-7B, a 7B parameter reasoning specialized model that matches or exceeds many 14B…
TII Abu-Dhabi Released Falcon H1R-7B: A New Reasoning Model Outperforming Others in Math and Coding with only 7B Params with 256k Context Window
Technology Innovation Institute (TII), Abu Dhabi, has released Falcon-H1R-7B, a 7B parameter reasoning specialized model that matches or exceeds many 14B to 47B reasoning models in math, code and general benchmarks, while staying compact and efficient. It builds on Falcon H1 7B Base and is available on Hugging Face under the Falcon-H1R collection. Falcon-H1R-7B is interesting because it combines 3 design choices in 1 system, a hybrid Transformer along with Mamba2 backbone, a very long context that reaches 256k tokens in standard vLLM deployments, and a training recipe that mixes supervised long form reasoning with reinforcement learning using GRPO.
nexttech-news.com
January 7, 2026 at 12:46 PM
2512.17351
言語モデルのアーキテクチャの違いを理解することは困難であり、特に学術規模の事前学習(例:13億パラメータ、1000億トークン)では、結果がノイズやランダム性に支配されることが多い。この課題を克服するため、我々は中核的なモデル能力を分離・評価する制御された合成事前学習タスクを導入する。この枠組...
January 8, 2026 at 12:11 AM
Gavin Tao, Yinuo Wang, Jinzhao Zhou
Can SSD-Mamba2 Unlock Reinforcement Learning for End-to-End Motion Control?
https://arxiv.org/abs/2509.07593
September 10, 2025 at 5:55 AM
Battery state of health estimation using a novel BiLSTM-Mamba2 network with differential voltage features and transfer learning
Battery state of health estimation using a novel BiLSTM-Mamba2 network with differential voltage features and transfer learning
Publication date: 28 February 2025Source: Journal of Energy Storage, Volume 110Author(s): Yunong Liu, Yuefeng Liu, Hongyu Shen, Liuxu Ding
www.sciencedirect.com
January 8, 2025 at 2:22 PM
experiment with Mamba2-2.7B, FastMamba attains 6\times higher energy efficiency than RTX 3090 GPU. [7/7 of https://arxiv.org/abs/2505.18975v1]
May 27, 2025 at 5:55 AM
complexity. Finally, we evaluate FastMamba on Xilinx VC709 FPGA. For the input prefill task on Mamba2-130M, FastMamba achieves 68.80\times and 8.90\times speedup over Intel Xeon 4210R CPU and NVIDIA RTX 3090 GPU, respectively. In the output decode [6/7 of https://arxiv.org/abs/2505.18975v1]
May 27, 2025 at 5:54 AM
a dedicated accelerator on FPGA with hardware-algorithm co-design to promote the deployment efficiency of Mamba2. Specifically, we successfully achieve 8-bit quantization for linear layers through Hadamard transformation to eliminate outliers. [3/7 of https://arxiv.org/abs/2505.18975v1]
May 27, 2025 at 5:54 AM
arXiv:2505.18975v1 Announce Type: new
Abstract: State Space Models (SSMs), like recent Mamba2, have achieved remarkable performance and received extensive attention. However, deploying Mamba2 on resource-constrained edge devices encounters many [1/7 of https://arxiv.org/abs/2505.18975v1]
May 27, 2025 at 5:54 AM
Another great hnet deep dive by @main_horse (on the other network). If you ever care about making tokenizerless happen and gigantic mamba2 layer. main-horse.github.io/hnet/eng-1gpu/
September 3, 2025 at 12:39 PM
🤖 Une IA qui conçoit ses propres architectures neuronales? Des chercheurs affirment avoir créé un système générant automatiquement 2000 nouveaux designs d'IA. Résultat: améliorations de 1-3% sur Mamba2. Les experts restent divisés - révolution ou simple optimisation? 🧠⚡ patb.ca/r/js0
July 31, 2025 at 10:27 PM
Euiyeon Kim, Yong-Hoon Choi: Mamba2 Meets Silence: Robust Vocal Source Separation for Sparse Regions https://arxiv.org/abs/2508.14556 https://arxiv.org/pdf/2508.14556 https://arxiv.org/html/2508.14556
August 21, 2025 at 6:34 AM
Zhi Qin Tan, Xiatian Zhu, Owen Addison, Yunpeng Li: U-Mamba2-SSL for Semi-Supervised Tooth and Pulp Segmentation in CBCT https://arxiv.org/abs/2509.20154 https://arxiv.org/pdf/2509.20154 https://arxiv.org/html/2509.20154
September 25, 2025 at 6:31 AM
Pengfei Hao, Hongqiu Wang, Shuaibo Li, Zhaohu Xing, Guang Yang, Kaishun Wu, Lei Zhu: Surgical-MambaLLM: Mamba2-enhanced Multimodal Large Language Model for VQLA in Robotic Surgery https://arxiv.org/abs/2509.16618 https://arxiv.org/pdf/2509.16618 https://arxiv.org/html/2509.16618
September 23, 2025 at 6:31 AM
Zhi Qin Tan, Xiatian Zhu, Owen Addison, Yunpeng Li: U-Mamba2: Scaling State Space Models for Dental Anatomy Segmentation in CBCT https://arxiv.org/abs/2509.12069 https://arxiv.org/pdf/2509.12069 https://arxiv.org/html/2509.12069
September 16, 2025 at 6:32 AM
Nan Mu, Ruiqi Song, Zhihui Xu, Jingfeng Jiang, Chen Zhao: SFD-Mamba2Net: Strcture-Guided Frequency-Enhanced Dual-Stream Mamba2 Network for Coronary Artery Segmentation https://arxiv.org/abs/2509.08934 https://arxiv.org/pdf/2509.08934 https://arxiv.org/html/2509.08934
September 12, 2025 at 6:30 AM
for inference of which we compared the methods of the top three performing teams on both the RIP tasks: one state-space model (Mamba2) and two learning-based approaches (SVM and CNN-LSTM). The results indicate that the state-space model outperformed [5/7 of https://arxiv.org/abs/2503.08437v1]
March 12, 2025 at 6:03 AM