Tag
The author's red-black tree implementation in a new language for Wasm GC matches native C in speed and memory usage, with plans to publish the codebase.
RadixArk/Qwen3.8-Flash-Next-NVFP4 is now supported in SGLang-V100, enabling full context operation on 4xV100 GPUs with performance metrics showing high throughput and context handling up to 256k tokens.
This article provides a deep-dive into Retrieval-Augmented Generation (RAG) and vector search, with measured benchmarks on 100,000 documents showing the trade-offs between exact search and IVF index for speed and recall.
An update on the Ornith-1.0-35B GGUF model introduces a native MTP speculative-decode graft for faster inference on a single GPU, achieving ~1.3-1.35x decode speedup while maintaining near-identical token distribution. Benchmark numbers for throughput, TTFT, and long-context performance across multiple quants are provided.
Chinese AI models like DeepSeek and Qwen deliver competitive performance at 5x–20x lower cost than Western counterparts, reshaping the economics of AI and driving multi-model deployment strategies.