SuperAPI AI Directory
Partner with us中文AI newsFavorites

AI developer platforms

3,662 results · Page 35 of 102

PhiZero

AI developer platforms

PhiZero is a world model built around a "physical language", proposed by Shuyao Shang, Yuqi Wang, Ruopeng Gao, Xu Chen, Tieniu Tan, Lue Fan, Zhaoxiang Zhang, and others. Its core idea is that the model first reasons about future dynamics using a discrete physical language, and then renders this reasoning result as video. This repository is the official implementation of PhiZero, providing training and inference pipelines, and stating that the code and checkpoints will be released soon. The repository includes training of the physical language tokenizer, physical language ground-truth export, training of the physical language reasoner, source-domain adaptation, and examples such as inference, motion transfer, and reconstruction. The project provides pretrained checkpoints and a unified Linux x86-64, CUDA 12.4 environment configuration.

Read more

picotron

AI developer platforms

Picotron is a minimalist 4D parallel distributed training framework released by Hugging Face, intended for educational purposes. Inspired by NanoGPT, it is designed specifically for pretraining Llama-like models and supports data parallelism, tensor parallelism, pipeline parallelism, and context parallelism. The code is concise and readable; train.py, model.py, and each parallelism module are all under 300 lines, making it suitable for learning and experimenting with distributed training techniques. Performance is still under active development; in measured tests, training LLaMA-2-7B on 64 H100s reaches 38% MFU, and training SmolLM-1.7B on 8 H100s approaches 50% MFU. The framework is mainly aimed at educational scenarios, helping users quickly become familiar with various distributed training techniques. It provides a tutorial video and code repository for building Picotron from scratch. Users need to obtain a token from Hugging Face to download models. It supports GPU and CPU environments, can generate configuration files via create_config.py, run locally with torchrun, or submit Slurm jobs via submit_slurm_jobs.py. Specific installation steps, complete dependencies, and detailed benchmarks are not yet provided.

Read more

PR1

AI developer platforms

Perception-R1 (PR1) is the official code implementation of the NeurIPS 2025 paper, proposing a scalable reinforcement learning framework that post-trains multimodal large language models (MLLMs) via Group Relative Policy Optimization (GRPO) to enhance visual perception capabilities. Its core innovations include: perceptual perplexity analysis, used to reveal the key threshold for effective reinforcement learning in perception tasks; rule-based reward shaping and GRPO optimization, enabling scalable policy learning; and significant improvements on multiple visual perception benchmarks, for example reaching 31.9% mAP on the COCO2017 validation set, becoming the first 3B-scale MLLM to achieve this performance. The project provides training and evaluation code, supports tasks such as grounding, counting, and detection, and users need to configure model and data paths themselves and separately download COCO images.

Read more

Practical Tips for Finetuning LLMs Using LoRA (Low-Rank Adaptation)

AI developer platforms

This article is Sebastian Raschka's practical tips for LoRA (Low-Rank Adaptation) fine-tuning of large language models, summarized based on hundreds of experiments. LoRA is a widely used technique for efficiently training custom LLMs and is suitable for those interested in open-source LLMs. The article shares key lessons and answers common questions, including consistency of training results, QLoRA trade-offs, optimizer selection, application of LoRA layers, hyperparameter tuning, and more, aiming to save fine-tuning time.

Read more

Private LLM

AI developer platforms

Private LLM is an AI chat app that can run large language models locally on iPhone, iPad, and Mac. It supports running models such as DeepSeek R1, Llama 3.3, Qwen3, and Gemma 3, is completely offline, does not censor content, and is integrated into Siri and Apple Shortcuts. It uses a one-time purchase, no-subscription model and supports Family Sharing for up to six people. Developed by two engineers, it uses OmniQuant and GPTQ quantization techniques to optimize output and reduce memory usage.

Read more