Libraries & frameworks
AngelSlim
Tencent/AngelSlim
Tencent Hunyuan's model compression toolkit for large models. It bundles quantization, speculative decoding, sparse attention, and distillation to make LLM deployment cheaper and faster.
Topics
Related projects
sglang38.7kA fast open-source inference framework for serving large language and multimodal models at scale.
peft21.9kA Python library for parameter-efficient fine-tuning like LoRA, adapting large pretrained models by training only small adapters.
tinyengram1.3kOpen research codebase reproducing DeepSeek-style Engram memory-injection architecture on Qwen-3 and Stable Diffusion, comparing against LoRA fine-tuning with training experiments and vision extensions.
loongforge832A high-performance training framework from Baidu for LLMs, VLMs, diffusion, and embodied models, supporting pretraining, SFT, LoRA, and distributed training on GPUs and Kunlun XPUs.
dsh-our-free-model72.7kThis DeepSeek Harness plugin adds free, no-key access to several upstream model pools and exposes a local OpenAI/Anthropic-compatible gateway. It includes fallback routing, usage dashboards, token-saving compression and in-app updates.
dsh-our-free-model58.9kA DeepSeek Harness plugin that adds free and bring-your-own-key model providers, a local OpenAI/Anthropic-compatible gateway, routing, usage dashboards, and token compression.