Independent coverage of frontier models, Indian compute initiatives, practical implementation guides, and open benchmarks.
How DeepSeek-V3 achieved 671B parameters with only 37B activated per token, training on just 2,048 H800 GPUs for under $6 million.
Meta's latest open-weights model achieves 405B-class benchmark performance on a single 8x H100 node or quantized 4-bit dual RTX 4090s.
Mistral's first multimodal model processes arbitrary aspect ratios without tiling artifacts. We tested it on Indian tax invoices and technical diagrams.
Claude 3.5 Sonnet can now interact directly with computer desktop GUIs, moving cursors, clicking buttons, and completing browser workflows.