A practical account of setting up an open-weight model stack for agentic software development in an enterprise context. The team evaluated top open-weight models (GLM-5.2, DeepSeek V4, Kimi K-2.6), noting that competitive options currently come exclusively from Chinese labs. Key considerations covered include model hosting via OpenRouter with a single vetted provider (filtering for quantization, latency, throughput, uptime, and data training policies), and coding agent selection (settling on OpenCode over Pi and Cline). The post highlights that hosting availability is the most constrained part of the stack for enterprise teams, and that model routing marketplaces may not suit most organisations' long-term governance requirements.
