xAI wrapped training on Grok V9-Medium last week — a 1.5 trillion parameter model, three times the size of the v8-small currently running in production. Supervised fine-tuning is underway with reinforcement learning next, putting a public release somewhere in mid-June 2026.
What the parameter jump actually means
Raw parameter count isn’t the whole story, but going from 500B to 1.5T with a specific focus on coding and agentic tasks is meaningful. The bet is that scale, combined with RL-heavy fine-tuning, closes the gap with GPT-5.4 and Claude 4 on multi-step tool use — which is where the current Grok generation falls short.
What to watch
xAI also confirmed the current v8-small model will be open-sourced by end of 2026. For builders, that’s the more immediately useful development — a production-grade 500B model you can self-host changes the cost equation for high-volume workloads.
Mid-June is close enough to start benchmarking your agent pipelines against it when it lands.