Tag
Liquid AI 推出新一代 Liquid Foundation Models 系列 LFM3，采用共享骨干的多模态（文本、视觉、音频、全模态、决策）混合架构，首批 2B 参数级模型开放权重，可本地微调并跨设备运行。
论文提出 Hybrid Reasoning Agents（HybridAgents），通过轻量级 Hybrid Link 在潜在推理与文本交互之间动态切换，结合 Hybrid SFT 与 Hybrid RL 后训练算法，在 11 个基准上较前沿 LLM 智能体平均提升 10.4%，同时实现 35.5%–54.7% 的 token 削减、1.8×–2.6× 推理加速和 2.7×–3.7× 成本节省。
Technology Innovation Institute introduces Falcon-Emirati-7B, a dialect-specialized LLM built on Falcon-H1-Arabic's hybrid Mamba-Transformer architecture, trained to understand and generate Emirati Arabic — including dialect vocabulary, tone, and cultural context like nabati poetry and proverbs — that general Modern Standard Arabic models miss.
Luke Salamone distilled Stockfish evaluations into 3.9 billion chess positions and trained a hybrid ResNet-Transformer value network on 1 billion of them, releasing the open Gigafish 3.8B state-value dataset on Hugging Face — reportedly the largest open chess state-value dataset to date.
This paper presents an exploratory ablation study of TALH, a hybrid language model combining MLA and SSM, showing that SSM integration is more critical for validation performance than MLA in the tested setup, with insights on memory usage and timing on consumer hardware.
The user explores adding AI object detection to older security cameras without straining network bandwidth, considering edge processing with on-premises appliances to analyze video locally and only send alerts or short clips, while seeking advice on hybrid architecture and API integration.
This paper proposes intervention-based metrics to differentiate retrieval and positional heads in RoPE Transformers, leading to a principled hybrid architecture (HwH) that combines full and linear attention for improved language modeling and long-context extrapolation.
InclusionAI introduces Ling-3.0-tiny, a 7.9B-parameter hybrid reasoning MoE model with only 1.3B active parameters per token, optimized for efficient local and edge deployment.
This paper proposes a hybrid Mamba-attention architecture for block diffusion language models that restricts reverse Mamba scans to the active denoising block, enabling exact caching across blocks and achieving high throughput for long-context generation.
Proposes Geometry-aware R-Structured KAN (GRS-KAN), a hybrid neural architecture that integrates R-functions into KAN to encode geometric and logical constraints, achieving up to 67% RMSE reduction on regression benchmarks with discontinuities.
Introduces Neuro-Bayesian-Symbolic Residual Attention Shallow Network (NBS-RASN), a hybrid neural architecture for explainable cybersecurity risk assessment in open-source ecosystems, using 80 interpretable neurons across 12 layers with hard constraints for interpretability.
NVIDIA's Nemotron-3-Super-120B-A12B, a hybrid Mamba and mixture-of-experts model, achieves perfect needle-in-haystack retrieval at 504K tokens using only four RTX 3090 GPUs.
The article argues that knowledge graphs and vector databases serve different purposes in enterprise AI and should be used together rather than as alternatives. It recommends hybrid architectures or managed solutions like 60x to handle both semantic recall and structural reasoning.
This paper presents two agentic AI frameworks, DeepTS/DeepCollector and DeepScribe, that automate scientific workflows including time-series data curation and conversion of physics lectures into structured reports, using a hybrid local-cloud architecture with LLMs.
Interfaze introduces a new hybrid AI model architecture that combines DNN/CNN encoders with transformers to achieve superior accuracy and cost-efficiency for deterministic tasks such as OCR, vision, and STT, compared to generalist models.
The article discusses the growing viability of local AI models for everyday tasks, suggesting a shift toward hybrid architectures that optimize for cost and latency rather than relying solely on frontier cloud models.