Command Palette
Search for a command to run...
128K Long Context + Agent Reinforcement Training! LFM2.5-2.6B Unlocks Efficient Deployment of Large Models on Edge; DETR Uses Transformer to Cut Through NMS and Anchors, Reshaping Object Detection
As AI agents gradually move from the cloud to the terminal, achieving high-quality task execution with limited computing power and memory becomes crucial. LFM2.5-2.6B is designed to meet this need. Based on the LFM2 architecture, it introduces 128K long context and agent reinforcement training, demonstrating superior performance in tool invocation, instruction compliance and multi-step tasks, and can compete with models with 4 times larger parameter sizes.The model is also trained and optimized for mainstream agent frameworks to improve compatibility in practical applications. More notably, it achieves inference speeds of up to 220 tok/s on an Apple M5 Max and 113 tok/s on an AMD Ryzen CPU, with a memory footprint of less than 2.5GB, striking a good balance between lightweight design and high performance, and providing a new option for edge agent deployment.
The HyperAI website now features "LFM2.5-2.6B: An Efficient Agent Inference Model for Edge Deployment." Give it a try!
Online use:https://go.hyper.ai/rbtpt
A quick overview of hyper.ai's official website updates from August 7th to August 13th:
* High-quality public datasets: 5
* A selection of high-quality tutorials: 6
* Popular encyclopedia entries: 5
Visit the official website:hyper.ai
Selected public datasets
1. GLM-5.2-Conversation Dialogue Trajectory Dataset
GLM-5.2-Conversation is a dialogue trajectory dataset designed to provide a large-scale Chinese and English question-and-answer corpus for supervised fine-tuning (SFT) and inference distillation of large conversational models. The dataset contains 50,000 dialogue trajectory samples generated by GLM-5.2 in High Inference mode. Each sample consists of a user question and a complete inference response from the model, covering topics such as programming, mathematics, physics, creative writing, and logical question answering.
Online use:https://go.hyper.ai/T2GCb
2. XYZ-Aquila SFT Agent Multi-round Trajectory Dataset
XYZ-Aquila SFT is a multi-turn trajectory dataset for search-oriented agents released by XYZAILab. It aims to provide high-quality bilingual multi-turn tool-call trajectory data for supervised fine-tuning (SFT) of search agents. The dataset contains 7,000 multi-turn search-oriented tool-use trajectories, including 5,000 English examples and 2,000 Chinese examples. This dataset is a sample of the supervised fine-tuning corpus used by the XYZ-Aquila-mini and XYZ-Aquila-pro models. Each sample fully records one agent interaction process, covering the search tool call, intermediate observations, and the generation of both Chinese and English responses.
Online use:https://go.hyper.ai/t7r54
3. Recursive-Task-Synthesis: Synthesizing Recursive Task Datasets
Recursive-Task-Synthesis is a dataset of synthetic recursive tasks specifically designed for reinforcement learning and capability evaluation of command-line AI agents. It aims to provide researchers with executable and validated real-world task instances. The dataset contains 37,484 validated command-line task instances, each accompanied by natural language instructions, reference answers, source data, and a pre-configured Docker runtime environment. The tasks cover multiple technical areas, including software development, system maintenance, and script writing.
Online use:https://go.hyper.ai/ZsQ2L
4. PROTECT-90 Power System Protection Fault Dataset
PROTECT-90 is a power system protection fault waveform dataset designed to provide standardized and reproducible benchmark data for power system protection and transient analysis research. This dataset contains 9,022 electromagnetic transient (EMT) simulation cases, all generated under a standard 90 kV double-circuit topology. Each segment provides 1 second of synchronous three-phase voltage and current waveforms at a sampling rate of 6.4 kHz (128 sampling points per 50 Hz cycle), recorded at 8 protection-related measurement locations, with a total of 48 synchronous waveform channels.
Online use:https://go.hyper.ai/epqXm
5. Maritime Continent Satellite Weather Dataset
Maritime Continent Satellite Weather is a satellite meteorological imagery dataset for maritime continental regions, designed to explore the application of deep learning in weather forecasting and nowcasting, and to support research in rainfall prediction, thunderstorm identification, and multimodal model fusion. The dataset contains 95,036 hourly acquired Himawari-8/9 satellite infrared image frames paired with meteorological data, covering the western maritime continental region (Southeast Asia and the Strait of Malacca, covering Singapore, Indonesia, Malaysia, and Thailand), and is accompanied by surface weather data for 23 cities in these countries. Each city file contains 97,056 rows of data.
Online use:https://go.hyper.ai/2dOI3
Selected Public Tutorials
1. LFM2.5-2.6B: Efficient Agent Inference Model for End-User Deployment
Released by Liquid AI in August 2026, LFM2.5-2.6B is a lightweight hybrid architecture language model in the LFM2.5 series designed for edge deployment. It supports 128K ultra-long contexts and possesses strong agent tool invocation capabilities. The model employs 22 dual-gated short convolutional modules and 8 GQA modules, achieving a balance between performance and computational efficiency. It is post-trained using agent reinforcement learning and can be used in scenarios such as agent workflow, long document understanding, RAG, and edge inference.
Run online:https://go.hyper.ai/rbtpt
2. NLP Tutorial: Patent Phrase Similarity Matching Based on DeBERTa
This tutorial is designed for NLP beginners, using patent phrase similarity matching as a case study to explain the core processes of natural language processing from scratch. Based on Microsoft DeBERTa-v3 and HuggingFace Transformers, the tutorial covers data exploration, word segmentation, model training and evaluation, and can be used for tasks such as patent text similarity analysis, semantic search, text classification and regression.
Run online:https://go.hyper.ai/3R3Yp
3. DETR: End-to-end object detection using Transformer
DETR (DEtection Transformer), released by Facebook AI Research (FAIR) in 2020, is an end-to-end object detection model that redefines object detection as a direct set prediction problem. It replaces the complex post-processing steps such as NMS and Anchors in traditional detectors with a Transformer architecture. The model employs an Encoder-Decoder structure and Object Queries, and uses the Hungarian algorithm to achieve bipartite graph matching between the predicted results and the real objects. It can be used for object detection, as well as visual tasks such as instance segmentation and panoptic segmentation.
Run online:https://go.hyper.ai/eVfK1
4. GAN-BERT: Semi-supervised text classification based on generative adversarial networks
GAN-BERT was released in July 2020 by Tor Vergata's team at the University of Rome Tor Vergata in collaboration with Amazon. It introduces a Generative Adversarial Network (GAN) into a fine-tuned version of BERT, enabling efficient text classification with only a small amount of labeled data through collaborative training of the generator and discriminator. In the TREC fine-grained problem classification task, it achieved an accuracy of 41.81 TP3T using only 109 labeled samples, significantly outperforming the traditional BERT's 13.61 TP3T.
Run online:https://go.hyper.ai/ujIry
5. Inflect-Micro-v2: TTS with parameters below 10M
Inflect-Micro-v2, released by Owen Song in April 2026, is a lightweight end-to-end text-to-speech (TTS) model containing only about 9.357 million deployable parameters. The FP32 model is only 37.53 MB in size and supports CPU and CUDA inference. Based on the VITS architecture, the model can directly generate 24 kHz mono audio without an additional vocoder and supports long text segmentation, pause control, and deterministic seed reproduction. It is suitable for scenarios such as local speech synthesis, edge device deployment, and text reading.
Run online:https://go.hyper.ai/oAocJ
6. DeePMD-kit: A Tutorial on Deep Potential Molecular Dynamics
DeePMD-kit, first released by the DeepModeling community in 2018, is a deep learning-based molecular dynamics simulation toolkit. It uses neural networks to learn the interatomic potential surface in ab initio computational data, achieving large-scale molecular simulations with near-first-principles accuracy. This tool supports multiple deep learning frameworks such as TensorFlow, PyTorch, JAX, and PaddlePaddle, and can be integrated with mainstream molecular dynamics software such as LAMMPS and GROMACS. It won the 2020 ACM Gordon Bell Prize for its high-performance computing applications. It is widely used in fields such as material structure prediction, chemical reaction pathway exploration, catalyst design, and protein interaction simulation.
Run online:https://go.hyper.ai/F0H9W
💡We have also established a Stable Diffusion tutorial exchange group. Welcome friends to scan the QR code and remark [SD tutorial] to join the group to discuss various technical issues and share application results~
Popular Encyclopedia Articles
1. Optical Character Recognition (OCR)
2. World Action Model WAM
3. Remote Sensing
4. Glitch Token (a term used to describe a glitch-related term)
5. Generative Pre-trained Transformation Model (GPT)
Here are hundreds of AI-related terms compiled to help you understand "artificial intelligence" here:
The above is all the content of this week’s editor’s selection. If you have resources that you want to include on the hyper.ai official website, you are also welcome to leave a message or submit an article to tell us!
See you next week!
