yesterday, elon musk officially confirmed that xai’s new flagship large model, grok v9-medium, has completed its foundational training and has officially entered the critical supervised fine-tuning phase. with 1.5 trillion parameters—three times the size of the current mainline v8-small version (0.5 trillion parameters)—this model not only achieves a significant leap in scale but also makes groundbreaking advances in logical reasoning depth, multi-step task decomposition, and long-range contextual modeling capabilities.
notably, during the post-training phase, v9-medium has, for the first time, systematically integrated high-quality real-world programming conversation data provided by the cursor platform—covering millions of lines of code written in real-time by engineers within ides, along with debugging notes, refactoring intentions, and interactive problem-solving processes. this move directly addresses the core bottleneck exposed by v8-small: insufficient density of professional engineering practices in the training corpus and a disconnect between code semantics and development intent. by shifting to a paradigm of “learning from developer behavior,” xai ensures that the model no longer merely studies static code snippets but instead internalizes the thought processes and collaborative logic embedded in actual coding scenarios, aiming to build an embodied ai engineer capable of understanding code at the repository level and autonomously planning and executing complex software tasks.
currently, the model is undergoing intensive human‑feedback reinforcement learning (rlhf) preparation, with full-scale evaluations expected to begin in 2–3 weeks. the first batch of access is scheduled to open to x platform developers and tesla’s in‑vehicle ai ecosystem around mid-june. crucially, v9-medium is natively designed for nvidia’s blackwell architecture and has undergone deep co‑optimization in fp4 quantization and moe dynamic sparse activation, resulting in measured inference throughput improvements of over 2.3 times and a nearly 40% reduction in per‑token cost. this technological roadmap marks xai’s transition from general language modeling toward vertically specialized engineering intelligence infrastructure, and its deployment pace and performance are likely to reshape the competitive landscape of the ai programming assistant market.
