Grok Survives xAI Shutdown with 1.5‑T V9‑Medium Model – Musk Announces

Following the disbanding of xAI and the departure of many founding members, the community wondered whether Grok would be discontinued. Musk confirmed that Grok is still advancing, announcing that the 1.5‑trillion‑parameter Grok V9‑Medium foundation model has finished training and achieved promising evaluation results.

The model is now entering a fine‑tuning stage and will start a reinforcement‑learning phase in a few days. According to the roadmap, Grok V9‑Medium is expected to be publicly released within two to three weeks.

Musk highlighted the shortcomings of the current production model, Grok v8‑small (0.5 T parameters, external test version Grok 4.2), which suffers from serious deficiencies in training‑data quality, coverage, and proportion. V9‑Medium triples the parameter count, is specially optimized for Blackwell‑architecture GPUs, and is expected to deliver noticeable gains across all capabilities, especially in complex programming tasks.

The 0.5 T model is planned to be open‑sourced by the end of the year, enabling developers to run, fine‑tune, or build applications locally.

In the supplemental training phase, the team incorporated a large amount of data from the AI‑code editor Cursor, a data source that will continue to expand. This inclusion follows Musk’s $60 billion acquisition of Cursor announced by SpaceX in April, a move that sparked debate about the editor’s valuation but was defended on the grounds that Cursor captures real developer workflows (coding, debugging, iteration, bug fixing) and thus provides unique insight into how human developers think.

Beyond the new model, Musk recently launched Grok Build, an AI coding agent that runs directly in the command line. It functions like a Musk‑branded “Claude Code,” capable of reading code, planning, modifying files, and executing tests.

Grok Build’s standout feature is Plan Mode : after a user describes a requirement, the agent first outputs a structured execution plan, listing each step. Users can review, edit, or rewrite the plan, and only after approval does the agent apply changes, presenting modifications as diffs.

Additional key capabilities include up to eight parallel sub‑agents, an Arena Mode that pits multiple candidate solutions against each other, native support for MCP and ACP, and headless operation.

In the terminal‑native programming‑agent arena, Grok Build joins three existing contenders: Anthropic’s Claude Code, OpenAI’s Codex CLI, and Google’s Gemini‑powered Antigravity, turning a three‑horse race into a four‑horse competition.

Musk previously admitted that xAI lagged behind in coding ability, positioning Grok Build as a “catch‑up” entry that quickly matches rivals in architecture and then attempts to overtake them with differentiators like Arena Mode.

However, Grok Build currently runs on the grok‑code‑fast‑1 engine with a 256 K‑token context window, which is modest compared with competitors and may become a bottleneck for large codebases or long‑running agent sessions.

The 1.5 T parameters of Grok V9‑Medium, combined with the injection of authentic Cursor workflow data, are expected to strengthen the underlying model for the entire toolset.

Whether Musk’s team can secure a lasting position in the emerging AI‑agent battlefield remains to be seen.