{"id":85205,"date":"2026-06-25T01:48:14","date_gmt":"2026-06-25T01:48:14","guid":{"rendered":"https:\/\/www.europesays.com\/ai\/85205\/"},"modified":"2026-06-25T01:48:14","modified_gmt":"2026-06-25T01:48:14","slug":"cursor-trains-first-frontier-model-from-scratch-on-colossus-1-5-trillion-parameters","status":"publish","type":"post","link":"https:\/\/www.europesays.com\/ai\/85205\/","title":{"rendered":"Cursor Trains First Frontier Model From Scratch on Colossus: 1.5 Trillion Parameters"},"content":{"rendered":"

At its Compile keynote<\/a> on June 16, Cursor co-founder Michael Truell disclosed that the company is training its first fully self-built AI model \u2014 a 1.5-trillion-parameter system pre-trained from scratch on xAI’s Colossus supercomputer cluster in Memphis, Tennessee \u2014 and expects to ship it to users within weeks. The announcement, made the same day SpaceX confirmed its $60 billion all-stock acquisition of Cursor parent Anysphere, marks the first time the AI coding tool has attempted to build a model at the foundational level rather than adapting one that someone else built first.<\/p>\n

The distinction matters architecturally. Every prior Cursor model, including Composer 2 and Composer 2.5, was built by taking a published open-source foundation \u2014 specifically the Kimi K2.5 base from Moonshot AI \u2014 and performing continued pretraining and reinforcement learning on top of it. That approach meant Cursor inherited the architectural choices embedded in someone else’s pre-training: what data the model saw, how its parameters were initialized, what conceptual structures its weights encoded before Cursor touched it. The new model begins from a blank initialization. Cursor now owns the full training pipeline, which means it controls not just how the model is fine-tuned for coding tasks but what the model fundamentally knows before any task-specific training begins.<\/p>\n

What Training From Scratch on 100,000 GPUs Actually Means<\/p>\n

The model is being pre-trained on more than 100,000 Nvidia GPUs across Colossus \u2014 the xAI supercomputer cluster in Memphis that launched in 2024 with approximately 100,000 H100-equivalent units and has since expanded to around 200,000. Training runs at this scale typically require datasets reaching trillions of tokens and several weeks of continuous computation, with gradients flowing across a tightly interconnected GPU fabric. The compute allocated to this training run is ten to twenty times greater than anything Cursor has used on prior models.<\/p>\n

Truell positioned the new model as comparable in size to Anthropic’s Claude Opus and OpenAI’s GPT-5.x tier \u2014 and confirmed, in a notably candid disclosure, that both of those model families currently fall below two trillion parameters. The observation reframes how the model-size competition works at the frontier: the gap between a well-resourced application-layer company with access to Colossus and a dedicated AI lab is now, by this measure, smaller than it has ever been.<\/p>\n

The economics driving this decision are straightforward. Cursor’s core business had been structured as an API reseller: it charged developers subscription fees and paid Anthropic, OpenAI, and other model providers for the inference that powered those subscriptions. The margin between what Cursor charged and what it paid was consistently squeezed, creating a ceiling on profitability that no amount of product growth could fully eliminate. A proprietary model trained on Cursor’s own developer workflow data removes that ceiling. If the model performs, inference costs drop from a per-token payment to a capex sunk into training \u2014 a fundamentally different cost structure.<\/p>\n

The model is also explicitly positioned as general-purpose, not coding-specific. Truell described it as designed to go beyond coding, an indication that Cursor sees this as a competitor to foundation models broadly rather than a narrowly specialized coding tool.<\/p>\n

Origin: Agent-First Git Hosting Built to Handle What GitHub Cannot<\/p>\n

The second announcement from Compile is Origin, a Git hosting and code collaboration platform that Cursor will open for broad availability in fall 2026. A waitlist is currently open at cursor.com.<\/p>\n

Origin was built on the technical foundation of Graphite<\/a>, the New York-based code review startup Cursor acquired in December 2025. Graphite co-founder Tomas Reimers, who joined Cursor through the acquisition, led the Origin demo at Compile, where the team showed throughput numbers that make the architectural premise explicit: approximately 296,000 repository clones per hour, roughly 81,000 pushes per hour, and 22.6 commits per second against a single repository, with global synchronization latency under 400 milliseconds and automatic failover under 10 milliseconds. The storage architecture uses NVMe-backed Git fileservers as the hot tier, with Amazon S3 as the source of truth for unlimited replica scaling.<\/p>\n

Those numbers are demo metrics rather than independently verified production benchmarks, and Cursor has not yet released pricing or detailed documentation. But the engineering constraint they describe is real and increasingly pressing. GitHub currently processes roughly 275 million AI agent commits per week \u2014 approximately fourteen times the platform’s total commit volume for all of 2025 combined \u2014 and GitHub’s infrastructure, designed for human-paced development workflows, has struggled to keep up. The platform experienced nine outages in May 2026 and has been running below its enterprise service-level agreement of 99.9% uptime. GitHub moved parts of its infrastructure onto AWS rather than Microsoft’s Azure in an attempt to absorb the load, a detail that underscores how much the agentic coding wave has strained a platform built for a different era.<\/p>\n

Origin’s architectural bet is that the problem is not resolvable by scaling GitHub’s existing design. When AI agents write code in parallel \u2014 dozens of simultaneous sessions all cloning the same repository, creating branches, making commits, and generating pull requests that conflict with each other \u2014 the merge conflict model that GitHub inherited from human-paced software development becomes a bottleneck at machine speed. Origin addresses this by building merge conflict resolution around semantic intent rather than line-level diff markers. Rather than flagging conflicting lines for a human to resolve, Origin’s conflict engine is designed to reason about what each agent’s branch was trying to accomplish, and to resolve the conflict based on that inferred intent. The diff format, review interface, and merge state logic are all designed to be machine-readable and machine-actionable from the start \u2014 not patched onto a human-centric design after the fact.<\/p>\n

For developers who are already uncomfortable handing proprietary code to a Cursor-owned hosting platform now under SpaceX’s pending acquisition, the data handling question remains unanswered. Cursor has not published Origin’s data retention terms. Until those terms are available, the same practical caution that applies to Cursor’s editor \u2014 reviewing Privacy Mode settings and understanding what code is transmitted and stored \u2014 applies to Origin as well.<\/p>\n

Cursor Mobile Puts Agent Control in Your Pocket<\/p>\n

The third product from Compile is Cursor Mobile, now available as an iOS public beta via TestFlight. The app is designed not as a coding environment for mobile development but as a remote management layer for AI agents: users can check on running agents, unblock tasks that have stalled waiting for input, and view and comment on screenshots that agents have generated during their work. A remote access feature allows users to connect to agents running on a local machine from their phone.<\/p>\n

The positioning reflects a broader shift in how Cursor defines its product. The editor is no longer the whole product \u2014 it is the execution environment. Cursor Mobile assumes the developer has delegated work to agents and needs a way to supervise that delegation from anywhere, not just from the desk where the IDE is open.<\/p>\n

What Developers Should Weigh Before the Model Ships<\/p>\n

The three Compile announcements taken together describe a company that is attempting to move from the application layer of the AI stack all the way down to its foundation \u2014 owning the pre-training architecture, the version control infrastructure, and the mobile supervision interface in addition to the code editor itself.<\/p>\n

That vertical integration carries real competitive advantages. A proprietary model trained on Cursor’s own developer workflow data creates a training flywheel that API-only competitors cannot replicate: more usage generates more proprietary training signal, which improves the model, which attracts more usage. It is the same data moat that GitHub’s twenty-six million users give Microsoft’s Copilot team, built through a different mechanism.<\/p>\n

It also carries real risks for developers. Cursor has historically offered model agnosticism as a core feature \u2014 the ability to route work to Anthropic’s Claude, OpenAI’s GPT, or Cursor’s own Composer depending on the task. The open question, flagged by analyst Jason Andersen of Moor Insights and Strategy, is whether Cursor under SpaceX ownership will continue to let users point to Grok alternatives. No public commitment to maintaining that model access has been made, and SpaceX has a direct financial incentive to increase Grok utilization after xAI’s $6.35 billion operating loss in 2025. Developers with enterprise contracts built around Claude access inside Cursor should begin watching SpaceX’s post-merger communications closely.<\/p>\n

Whether Cursor’s self-trained model can match the quality of dedicated AI labs at launch remains unknown \u2014 no benchmarks are available before the model ships. But if it performs as positioned, it puts Cursor in a category occupied by a very short list of companies: those that build the developer tool, pre-train the model from scratch, own the compute infrastructure, and control the hosting layer for the code their agents produce.<\/p>\n

Frequently Asked Questions<\/p>\n

What is Cursor’s new AI model and when does it ship?<\/p>\n

Cursor’s new model has more than 1.5 trillion parameters and is being pre-trained from scratch on xAI’s Colossus supercomputer cluster using more than 100,000 GPUs. Unlike prior Cursor models, which built on the open-source Kimi K2.5 foundation, this one begins from a blank initialization \u2014 giving Cursor control over the entire training pipeline. Cursor co-founder Michael Truell said at the June 16 Compile keynote that it was expected to ship within a few weeks of that date.<\/p>\n

What is Cursor Origin and how does it differ from GitHub?<\/p>\n

Origin is Cursor’s agent-first Git hosting platform, currently in waitlist preview with general availability targeted for fall 2026. Its central technical distinction from GitHub is that it was built from the ground up assuming AI agents \u2014 rather than human developers \u2014 will be the primary traffic source. The platform uses an NVMe-plus-S3 storage architecture with AI-powered merge conflict resolution designed to handle parallel agent commits at a scale GitHub’s human-centric design was not built for. Demo metrics showed roughly 22 commits per second and 296,000 clones per hour in a single repository.<\/p>\n

Will Cursor still support Claude and OpenAI models after the SpaceX acquisition closes?<\/p>\n

No changes to Cursor’s model access have been announced as of the acquisition filing on June 16, 2026. Cursor continues to support Anthropic’s Claude models, OpenAI’s GPT models, and its own Composer models. After the deal closes \u2014 expected in the third quarter of 2026 \u2014 SpaceX has a financial incentive to prioritize Grok, given xAI’s losses in 2025. No public commitment to maintaining multi-model access has been made. Developers who depend on routing work to Anthropic or OpenAI through Cursor should monitor SpaceX’s communications as the merger close approaches.<\/p>\n

What is Cursor Mobile and what can it do?<\/p>\n

Cursor Mobile is an iOS public beta available now through TestFlight. It is not a mobile code editor \u2014 it is a remote control interface for AI agents running in Cursor. Users can monitor running agents, unblock stalled tasks, review agent-generated screenshots, and access agents running on local machines via a remote connection feature.<\/p>\n","protected":false},"excerpt":{"rendered":"At its Compile keynote on June 16, Cursor co-founder Michael Truell disclosed that the company is training its…\n","protected":false},"author":2,"featured_media":85206,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[10],"tags":[11960,6541,45818,45819,8168,29187,658,2899,45820],"class_list":["post-85205","post","type-post","status-publish","format-standard","has-post-thumbnail","category-xai","tag-ai-coding-assistant","tag-cursor","tag-cursor-ai-model","tag-cursor-origin","tag-developer-tools","tag-frontier-ai-model","tag-spacex","tag-xai","tag-xai-colossus"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/85205","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=85205"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/85205\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/85206"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=85205"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=85205"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=85205"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}