Analysis of GLM-5.3-Flash architecture, a new LLM with hybrid attention and sparse MoE, compared to previous versions.
Analysis of Kimi K3's open-weight architecture, focusing on LatentMoE, attention residuals, and NoPE innovations for inference efficiency.
Mira Murati's Thinking Machines Lab releases Inkling, an open-weights multimodal AI model with 975B parameters, Apache-2.0 licensed.
Mira Murati's Thinking Machines Lab releases Inkling, an open-weights multimodal AI model with 975B parameters, Apache-2.0 licensed.
Analysis of Thinking Machines Lab's new open-weight 975B MoE LLM, Inkling, with benchmark comparisons and architectural surprises.
Analysis of Thinking Machines Lab's new open-weight 975B MoE LLM, Inkling, including benchmarks, architecture details, and comparisons.
Tencent releases Hy3, a 295B-parameter MoE model outperforming larger open-source models, available free on OpenRouter.
Tencent releases Hy3, a 295B-parameter MoE model outperforming larger open-source models, available for free on OpenRouter.
Analysis of Cohere's new North Mini Code model for agentic coding tasks, including architecture and benchmark performance.
Cohere's North Mini Code is a 30B-parameter open-weight model for agentic coding, featuring Mixture-of-Experts architecture and strong benchmark performance.
Explores Mixture of Experts (MoE) in AI models, its sparse routing principle, and how it enables large model capacity with low compute cost per token.
DeepSeek V4 preview models offer frontier-level performance at a fraction of the cost, with up to 1M token context and open weights.
DeepSeek releases V4 Pro and Flash AI models, offering frontier-level performance at significantly lower costs.
Explores using Apple's 'LLM in a Flash' research to run a massive 397B parameter AI model locally on a MacBook by streaming weights from SSD.
Explores using Apple's 'LLM in a Flash' research to run a massive 397B parameter AI model locally on a MacBook by streaming weights from SSD.
Mistral AI releases Mistral Small 4, a new 119B parameter open model combining reasoning, multimodal, and coding capabilities.
Mistral AI releases Mistral Small 4, a new 119B parameter open model combining reasoning, multimodal, and coding capabilities.
Analysis of NVIDIA's Nemotron 3 Super 120B-A12B model focusing on its accuracy-throughput trade-off design and efficiency features.
Analysis of NVIDIA's Nemotron 3 Super 120B-A12B model focusing on accuracy-throughput trade-off with Mamba-2 and MoE layers.
A hands-on guide to understanding and implementing the Qwen3 large language model architecture from scratch using pure PyTorch.
