More exciting open source projects: 1. OpenViking (ByteDance) https://((((((github.com))))))/volcengine/OpenViking ⭐ 12.8k | This week+6.5k A context database designed specifically for AI agents. Not a vector database, not a RAG framework. What it does is to manage all the things required by the Agent during runtime - memory, resources, skills - in a file system manner, deliver them to the Agent in layers as needed, and also self evolve. The first line of the readme reads' for AI Agents such as OpenClaw '. ByteDance was produced, and this week it entered the daily list and weekly list consecutively. 2. hermes-agent https://((((((github.com))))))/NousResearch/hermes-agent ⭐ 7.6k | This week+5.1k An agent that can evolve on its own. The core design has only one sentence: after completing complex tasks, automatically write this experience into a new skill; Next time you encounter a similar task, call this Skill and continue to improve it while using it. The smarter you run, the more skills you accumulate. $5 VPS can run. Telegram/Discord/WhatsApp/Signal/CLI support across all platforms. The model can be changed at any time, OpenRouter/Kimi/MiniMax/OpenAI are all good, switch with one command without changing the code. 3. BitNet (Microsoft) https://((((((github.com))))))/microsoft/BitNet ⭐ 34.7k | This week+6k The official reasoning framework for 1-bit LLM. What happens after quantization to 1-bit: ARM CPU acceleration 1.4-5x X86 CPU acceleration 2.4-6.2x • Energy consumption reduced by 55-82% A model with 100B parameters, where a single CPU can run No GPU required. The VPS you are currently using is sufficient. The matter of running large models locally is shifting from being a "game for people with GPUs" to something that everyone can play. 4. Deer flow (ByteDance) https://((((((github.com))))))/bytedance/deer-flow ⭐ 30.9k | This week+5k ByteDance's own SuperAgent framework has recently been open sourced. Can research, write code, and create content. Built in sandbox, memory, tools Skills、 Sub agents handle tasks ranging from minutes to hours. 5. nanochat https://((((((github.com))))))/karpathy/nanochat ⭐ 48.8k In 2019, OpenAI spent $43000 to train GPT-2. Karpathy uses this framework to run 8 H100 for 2 hours at a cost of $48, with comparable capabilities. Not a tool for ordinary users. It is designed for those who want to truly understand LLM - tokenization, pretraining, fine-tuning, inference, chat UI full process, with minimal code and easy modifications. 6. learn-claude-code（shareAI-lab） https://((((((github.com))))))/shareAI-lab/learn-claude-code ⭐ 28.3k | Today+870 "Bash is all you need." Peel off the core logic of Claude Code and re implement it from scratch using the simplest TypeScript+Bash, explaining every step clearly.