Tag
Release of Ling-3.0-tiny, a hybrid reasoning model with 7.9B total parameters and only 1.3B active per token, free for a week.
The builder of LOLM announces a hybrid Transformer-SSM language model and agent system from Qira, featuring a controller for live decisions, run receipts, a CLI, coding sandbox, MCP support, and lower-cost hosted access.
Liquid AI released LFM2.5-2.6B, a 2.6B-parameter hybrid model optimized for on-device deployment with 128K context, agentic post-training, and fast inference (220 tok/s on Apple M5 Max) under 2.5GB memory.
This paper presents a hybrid probabilistic forecasting system that integrates time series decomposition (Prophet) with NLP techniques applied to Bolivian news coverage to predict roadblocks, achieving improved AUC-ROC and Brier Score over purely statistical models.
This paper presents ClickGuard, a browser extension that detects and spoils clickbait news using a hybrid machine learning model combining transformer embeddings and linguistic features, achieving 91% F1-score on a combined dataset.
Cactus Hybrid is a post-trained Gemma 4 model that outputs confidence scores, allowing on-device inference with routing to larger models when confidence is low, achieving performance comparable to Gemini 3.1 Flash-Lite with minimal calls to the larger model.
This paper presents a hybrid approach for detecting online polarization in English and Hausa using DeBERTa for English and AfroXLMR-Social for Hausa and fine-grained subtasks, with LoRA and data augmentation to address computational and data constraints.
MiniCPM-SALA is a 9B-parameter hybrid attention model that interleaves sparse and linear attention to overcome the quadratic compute and large KV cache bottlenecks of long-context Transformers. It achieves 3.5x faster inference than Qwen3-8B at 256K tokens and supports up to 1M tokens on consumer GPUs, with a cost-effective continual training approach that reduces training costs by ~75%.
Graham Neubig shares a sidekick architecture for reducing LLM costs by delegating simple tasks to a smaller agent, with a 200-line example using the OpenHands SDK. This approach is also used in Cognition's Devin Fusion hybrid-model harness.
Proposes HybridCodec, a novel framework combining temporally compressed discrete tokens with continuous residuals to improve speaker characteristic retention in speech language models, reducing autoregressive steps while maintaining quality.
Liquid AI released LFM2.5-230M, a compact 230M-parameter hybrid model optimized for on-device deployment with fast edge inference speeds (213 tok/s on Galaxy S25 Ultra) and built for agentic tasks via reinforcement learning.
The NPU on AMD Strix Halo devices is now usable for AI inference, enabling hybrid mode that combines NPU and iGPU for faster prompt processing. Tools like Lemonade and AMD's ROCm software make this possible.
This paper presents a Hybrid NARX-LLM framework for predicting Greenland iceberg discharge, using a Physics-Informed Prompt method to guide an LLM for residual correction, improving accuracy over traditional NARX models.
D2H-AD is a novel anomaly detection framework using Hyperdimensional Computing (HDC) that combines distance-based and density-aware encoding. It outperforms five baselines across multiple benchmarks, offering lightweight, interpretable, and efficient performance for edge AI and IoT.
This paper proposes a hybrid classical-quantum variational autoencoder for neural topic modeling, embedding parameterized quantum circuits in the inference network. Experiments on the AgNews dataset demonstrate improved topic coherence and diversity compared to state-of-the-art classical models, showing viability on NISQ-era quantum devices.
AI voice agents like LuMay Voice Agent and Voxentis.ai are accelerating the replacement of traditional call centers due to advances in conversational AI, scalability, and performance consistency, driving hybrid AI-human models across industries such as healthcare, real estate, and e-commerce.
DAStatFormer is a hybrid multibranch Transformer that integrates statistical features with gated attention for efficient and accurate event classification in Distributed Acoustic Sensing (DAS), achieving up to 99.4% accuracy with significantly lower computational cost.
Introduces QuChaTeR, a hybrid architecture combining wavelet-based preprocessing, chaotic maps, and variational quantum circuits with recurrent structures for earthquake prediction, demonstrating faster convergence and superior accuracy over classical and quantum baselines.
This paper introduces TajPersLexon, a lexical resource for Tajik-Persian cross-script NLP, and benchmarks hybrid models against neural baselines to demonstrate effective low-resource processing.
Google announces Gemini 2.5 Flash, a new hybrid reasoning model available in preview through the Gemini API. The model features toggleable thinking capabilities, fine-grained thinking budgets for quality-cost-latency tradeoffs, and maintains fast inference speeds while improving performance over 2.0 Flash.