Releases·Tool Release
Baidu AI Cloud Baige team open-sources LoongForge training framework for LLMs, VLMs, diffusion and embodied models
Baidu AI Cloud's Baige team launched LoongForge, an open-source training framework on GitHub, announced via Show HN. It targets faster, cheaper training of mainstream LLMs, VLMs, diffusion, and embodied models, making it relevant to practitioners looking to cut training costs across modalities.
2 engagement1 sourceTue, Sep 15, 2026, 03:03 PM
AI Overview
- Baidu AI Cloud Baige team built LoongForge to speed training of mainstream LLMs, VLMs, diffusion and embodied models, cutting costs
- LoongForge is open-source with code hosted on GitHub
Related
Releases·Tool Release
LoongSage RL framework for LLMs released on GitHub
LoongSage, a production-grade reinforcement learning framework for LLMs, appeared on GitHub with validated training recipes. It targets high-efficiency RL post-training, giving practitioners a ready-to-use option for tuning large language models.
LoongSage, a production-grade reinforcement learning framework for LLMs, appeared on GitHub with validated training recipes. It targets high-efficiency RL post-training, giving practitioners a ready-to-use option for tuning large language models.
28 engagement1 sourcegithubFri, Aug 28, 2026, 07:22 AM
Releases·Tool Release
OpenArch – PyTorch implementations of modern LLM architectures
OpenArch: PyTorch implementations of modern LLM architectures released on GitHub
OpenArch: PyTorch implementations of modern LLM architectures released on GitHub
168 engagement2 sourceshackernews, redditMon, Sep 14, 2026, 07:55 AM
Releases·Tool Release
xllm: GitHub releases light-weight infrastructure for express LLM pre-training
xllm: lightweight infrastructure for express LLM pre-training
xllm: lightweight infrastructure for express LLM pre-training
29 engagement1 sourcegithubWed, Sep 2, 2026, 04:57 AM
Developments·Paper
Researchers introduce delta signal and EasyOPD framework to advance on-policy distillation for LLMs
Three papers from July 2026 collectively advance on-policy distillation (OPD) for LLMs. One introduces a 'delta signal' reward that compares teacher and base model outputs to improve token-level supervision. Another systematically studies OPD dynamics, finding it acts as an exploration catalyst and that prompt diversity matters more than per-problem sampling. The third presents EasyOPD, a unified framework built on verl to standardize fragmented OPD implementations, improving reproducibility and extensibility.
Three papers from July 2026 collectively advance on-policy distillation (OPD) for LLMs. One introduces a 'delta signal' reward that compares teacher and base model outputs to improve token-level supervision. Another systematically studies OPD dynamics, finding it acts as an exploration catalyst and that prompt diversity matters more than per-problem sampling. The third presents EasyOPD, a unified framework built on verl to standardize fragmented OPD implementations, improving reproducibility and extensibility.
Updated Thu, Jul 16, 2026, 04:07 PM — Delta signal reward introduced as alternative to direct teacher imitation in OPD.
0300 engagement3 sourcesarxivMon, Jul 13, 2026, 02:27 AM
Products·building
Horus-runtime: Train your own tiny LLM from scratch
Horus-runtime is a tool that lets you train a small language model from scratch, aimed at developers and researchers who want to understand LLM training or build custom small models. It provides a runtime environment for training, likely with a focus on simplicity and educational value.
Horus-runtime is a tool that lets you train a small language model from scratch, aimed at developers and researchers who want to understand LLM training or build custom small models. It provides a runtime environment for training, likely with a focus on simplicity and educational value.
1 engagement1 sourcehackernewsSat, Aug 1, 2026, 09:27 AM