I made all of you code products compete so fiercely that you forced Old Ma to step in again!
Musk posted on X that xAI's own Grok base model, V9-Medium (1.5T), has completed training.
We expect to officially launch in another 2 to 3 weeks:
Yes, this new model is clearly targeting programming skills with serious intent.
Musk specifically mentioned that the V9-Medium supplementary training incorporated a large amount of Cursor data, and more will be added in the future.
What does it mean——
(Probably meant to say: don’t just look at the other two when it comes to coding—take a closer look at mine too~)
The new model is being hyped up, and our in-house programming new product isn’t standing still either.
xAI’s newly released AI programming agent tool, Grok Build, is now in early Beta testing, and some users can already use it.
emm…
Less than a month after acquiring Cursor, Grok has already started enjoying a "bonus meal."
Feed new models with Cursor data while rolling out our own AI programming agent.
Anyway, Old Ma's code has arrived—better late than never!
The training of the new Grok model is complete!
Ultimately, coding is just too competitive...
From a timeline perspective, nearly a year has passed since the release of the Grok 4 series based on the v8-small base model.
Actually, the previous generation of models did see some improvement in programming capabilities, but it couldn't keep up with how quickly other companies were iterating!
So this time, xAI didn’t take small, incremental steps—they saved up a big move~
Grok's new base model, V9-Medium (1.5T), is directly targeted at programming capabilities and the developer market.
Even Musk himself can't wait to say: Programming skills have improved a lot. (doge)
In terms of parameters, the V9-Medium model directly reaches a scale of 1.5 trillion.
Currently, all production traffic for Grok is handled by a 0.5T v8-small model, which is also only 0.5T in size.
So, does that mean it only directly scales up to three times the parameters??
The significant increase in parameter scale also implies substantial improvements in the model's deep reasoning capabilities and knowledge base.
Greater model capacity often enables better understanding of complex contexts, more extensive chain-of-thought reasoning, and improved handling of challenging real-world development tasks.
However!
The most impressive aspect of this new model isn't in the parameters, but in its programming capabilities—
Ren Lao Ma specifically mentioned: xAI intentionally incorporated a large amount of 'Cursor data' during the training of V9-Medium (1.5T).
Everyone should know that Cursor's data isn't just a regular codebase—it's the real-world workflows of millions of developers.
Cursor data is more likely to show the model how developers in real projects describe requirements, identify issues, contextually guide AI, modify files, fix errors, and then follow up with further questions.
It’s also a solid dose of real-world engineering experience for Grok...
When will the model be released?
According to official statements, the base training for V9-Medium has been fully completed.
Fine-tuning is currently underway, and the reinforcement learning phase will begin in a few days.
According to xAI’s plan, the new model will be officially released to the public in another 2 to 3 weeks.
In addition, it is worth noting that—
While promoting his new model, Musk also left an easter egg in the comments:
The 0.5T model will be open-sourced by the end of this year—it’s also very practical!
(You can look forward to it a little bit~)
The AI programming tool Grok Build has also begun testing.
Beyond the new model, xAI's own AI programming tool has also made new progress.
Grok Build, an AI programming agent tool, is currently in early Beta testing.
Let me briefly introduce you: Grok Build is xAI’s own programming agent and CLI tool.
In form, it is more similar to terminal-based programming products like Claude Code and Codex CLI.
Developers can invoke it within their local project directory, allowing the AI to directly read the project, understand the context, plan changes, and modify code files, among other tasks.
Based on the official demo video, Grok Build’s standout capability is heavily focused on AI coding:
Plan Mode: After the user inputs their requirements, Grok Build does not proceed immediately but first creates a detailed execution plan for the user to review and approve before proceeding.
Multiple subagents work in parallel: For large, complex tasks, Grok Build assigns the work to subagents running in parallel.
Skills, Marketplace, MCPs Integration: We can easily extend Grok Build’s capabilities by invoking a variety of ready-to-use tools and plugins.
Additional capabilities: support for creating images and videos, building automated workflows, and more.
Regarding specific usage methods, the xAI official documentation also provides three entry points—
One is an interactive TUI.
This opens a full-screen interactive interface in the terminal, allowing developers to directly converse with Grok Build, view plans, and execute tasks.
One is headless mode.
Better suited for embedding in scripts, bots, or automated workflows to run tasks in the background according to instructions.
Another option is to connect to other applications via the Agent Client Protocol.
Convert Grok Build into a programming agent that can be invoked by external tools.
Many options available—no shortage of choices!
On one hand, xAI is enhancing the code capabilities of its next-generation foundational model using Cursor data.
On the other hand, they’ve launched Grok Build, a programming agent, directly entering the AI coding tools arena.
It's clear that Musk isn't just planning to compete with Grok in the chat box. (doge)
After acquiring Cursor, Musk rapidly jumped on the coding bandwagon.
Old Ma is again feeding new models with Cursor data and beginning testing on Grok Build.
It’s not really that sudden.
After all, several months ago, xAI began extensively engaging with and recruiting talent and expertise from Cursor while simultaneously undergoing internal layoffs and restructuring—
Andrew Milich and Jason Ginsberg, the two core engineering leads at Cursor who helped grow Cursor from zero to a valuation of billions of dollars.
Joined xAI in succession, reporting directly to Musk, with the core mission of rebuilding Grok’s coding capabilities from the ground up.
Then last month.
Elon Musk is no longer playing the game of poaching talent—he’s targeting Cursor itself.
Acquiring Cursor for $60 billion—this figure more than doubles Cursor’s valuation from last November…
So, looking back, just one month later, Musk's recent "inventory" of Cursor is already starting to show results on Grok:
Just half a month ago, Grok Build was only just emerging, and now it’s already being widely tested by developers.
On the V9-Medium side, Musk personally confirmed that a large amount of Cursor data has been incorporated into the supplemental training, with more additions planned for the future.
Moreover, during this time, Musk revealed that the new Grok model is being trained normally on the Colossus 2 cluster.
So many actions, so much code involved.
So, looking back at the phrase “more to come” in V9-Medium, it now seems quite intriguing.
I guess it translates to—
For each additional day Cursor takes on clients, Grok gets an extra meal.
Next, let’s see how Ma brings Grok into the coding arena—looking forward to it!
Reference link:
[1]https://openrouter.ai/x-ai/grok-build-0.1?utm
[2]https://x.com/elonmusk/status/2058787384364265734
[3]https://x.com/JasonBud/status/2058974659648123084
[4]https://x.com/xai/status/2058973760708091907
This article is from the WeChat public account "Quantum Bit," authored by Meng Yao.