GLM-5.1 (NovitaAI)

GLM-5.1 is Z.AI Zhipu's second-generation agentic-coding flagship, built around long-horizon autonomous engineering workflows that span hundreds of rounds and thousands of tool calls against full repositories. Its decoder draws on three independent design lineages: GLM-4.6 as the structural backbone, the MLA attention scheme popularized in DeepSeek V3 for compressing keys and values through low-rank latent projections, and the DeepSeek Sparse Attention indexer introduced in DeepSeek V3.2, which scores past tokens and retains only the top 2,048 keys per query so per-token attention cost stays roughly flat well beyond a few-thousand-token length. Together these choices support the model's roughly 200K-token working window, which is what makes sustained, multi-step repository editing practical rather than theoretical.

At the parameter level, GLM-5.1 is described as a sparse Mixture-of-Experts model on the order of 754B total parameters with roughly 40B active per token, trained on about 28.5 trillion tokens, and using a fine-grained 256-routed-expert plus one shared-expert layout with top-8 routing. The combination of Multi-head Latent Attention, the DSA indexer, and this MoE configuration is the practical reason the model is positioned for agentic coding and tool-heavy reasoning rather than pure chat. For teams evaluating open-weight deployments, this balance of a large total expert pool, a small active footprint, and sparse long-context attention tends to suit building autonomous development agents and research assistants that need to read large codebases while keeping per-call inference costs manageable on the offered gateway endpoints.