LumaByte

Microsoft's 41K-star BitNet project is now doing something particularly interesting for RAG and local AI. It has brought 1.58-bit weights to text embedding models. And in Microsoft's CPU benchmarks, its 600M-parameter embedding model achieves up to 2.28× the throughput of its FP16 baseline. This builds on the same BitNet research that produced the 2B-scale language model trained on 4 trillion tokens. The original idea was to train neural networks whose linear-layer weights take only three values: -1, 0, and +1. With specialized inference kernels, this can substantially reduce weight storage and memory traffic. Microsoft has now released two BitNet embedding models: → BitNet-embedding-270M, with 640-dimensional embeddings → BitNet-embedding-0.6B, with 1,024-dimensional embeddings Both support sequences up to 32,768 tokens and use ternary weights with 8-bit activations in their quantized linear layers. The 600M model is especially interesting. In Microsoft's reported English multilingual embedding benchmark comparison, it achieved an average score of 67.60, compared with 67.95 for its FP16 teacher. Meanwhile, with 8 CPU threads, the optimized 1.58-bit implementation delivered between 1.42× and 2.28× the FP16 prefill throughput, depending on sequence length. That performance difference matters for applications that repeatedly encode documents, queries, and other text. The implementation also reveals an important engineering detail. Its advantage becomes smaller as sequences grow longer, because attention computation starts accounting for more of the workload. And the project goes beyond embedding models. Microsoft has also released CPU optimization work, custom GPU inference kernels, and related low-bit speech-recognition infrastructure. The GitHub repository includes the inference engine, optimized kernels, supported models, and conversion tools. For engineers working on RAG, local inference, or resource-constrained AI systems, BitNet is a very interesting project to study. Fully open source under MIT. GitHub: github.com/microsoft/BitNet

Harness Engineering from Scratch 🚀 Found this excellent 21K+ star GitHub repo that teaches you how to build reliable AI coding agents, from the ground up. An entire open-source curriculum with 14 lectures, 8 hands-on projects, and 8 progressive learning phases. What makes it useful is the approach: instead of just learning concepts, you gradually build the engineering environment around an AI agent, using the same real application throughout the course. What's inside: → Agent instructions and repository structure → Context management and persistent memory → Multi-session continuity and task handoffs → Task planning, scope control, and progress tracking → Testing, verification, and observability → Automated coding loops and maker-checker workflows → Graph orchestration, parallel execution, and rollback Includes practical projects, reusable templates, a harness-creator skill, and technical breakdowns of how Claude Code, Codex, Pi, and DeepSeek build their harnesses. Available in 15 languages. Perfect for software engineers and AI developers who want to move beyond prompting and understand how reliable coding agents are actually built. Completely free and open source. Here's the GitHub Repo: github.com/walkinglabs/learn…