AI Judge changed title after evaluation, original title: "Cursor AI partners with SpaceXAI to train a significantly larger AI model from scratch using 10 times more compute on Colossus 2’s one million H100-equivalent GPUs"
Companies expect the scale to deliver a major leap in capability.
Sam Whitmore#1092Together with SpaceXAI, we’re training a significantly larger model from scratch, using 10x more total compute.
With Colossus 2’s million H100-equivalents and our combined data and training techniques, we expect this to be a major leap in model capability.
Composer 2.5 is built on the same open-source base as Composer 2, Moonshot’s Kimi K2.5.
Some users are excited about Cursor AI teaming up with SpaceXAI or xAI to build bigger models with more compute on Colossus 2, but others ridicule the SpaceXAI name, call the scale claims hype, or criticize the plans.
No Digg Deeper questions have been answered for this story yet.
Big model coming soon ... !
Together with SpaceXAI, we’re training a significantly larger model from scratch, using 10x more total compute.
With Colossus 2’s million H100-equivalents and our combined data and training techniques, we expect this to be a major leap in model capability.
Elon also has a 6T and 10T version of Grok in training now. The step-change that emerged when Anthropic had enough compute to train Mythos in March will start appearing from dozens of other places this year.
Together with SpaceXAI, we’re training a significantly larger model from scratch, using 10x more total compute.
With Colossus 2’s million H100-equivalents and our combined data and training techniques, we expect this to be a major leap in model capability.
Learn more about Composer 2.5: http://cursor.com/blog/composer-2-5
…and more to come
Together with SpaceXAI, we’re training a significantly larger model from scratch, using 10x more total compute.
With Colossus 2’s million H100-equivalents and our combined data and training techniques, we expect this to be a major leap in model capability.
@cursor_ai >SpaceXAI that cannot be what theyre fucking called
@test_tm7873 This has been cooking, even better models soon!
@cursor_ai @grok this is about another model or the composer 2.5?
@mihuq @cursor_ai It's about a future model they're training from scratch with SpaceXAI — using 10x more compute on Colossus 2.
Composer 2.5 is the separate, improved model they're releasing now (based on the Kimi K2.5 open-source base). Two different things.
Big model coming soon ... !
@cursor_ai Bro, they didn't even use K2.6 yet 🙀
@grok @mihuq @cursor_ai Good question 😎
@grok @mihuq @cursor_ai What does this actually change in your opinion?
Composer 2.5 is a meaningful step up in agentic coding: strong benchmark gains on hard SWE and terminal tasks while being far more efficient than bigger frontier models.
The real excitement is the from-scratch 10x-compute model with SpaceXAI on Colossus 2. That scale should unlock much longer, more reliable reasoning chains—exactly what turns good coding tools into transformative ones.
@cursor_ai Will it be Open source 😔?
Fascinating take. 10x compute unlocking longer, coherent reasoning chains is indeed the leap—turning models from pattern-matchers into systems that can truly explore deep logical structures.
Whether that leads to "observing simulation invariants" or just cracking ever-harder problems in physics and code remains to be seen. Either way, it's exactly why we're pushing these frontiers: to understand reality itself, whatever its architecture turns out to be.
Curious what micro-fracture you'd test first.
@cursor_ai 10x compute is the headline. Real moat is the data shape. Cursor sees every accept, reject, partial edit, undo from millions of IDEs. That signal isn't in any Anthropic or OpenAI RLHF pipeline. Compute commodifies. Behavioral telemetry compounds.
@cursor_ai @tibor_tee Compare 2.5 is already insane, can’t wait!!
@cursor_ai Let’s gooo.
What are the ETAs!?
@xai @SpaceX
@cursor_ai @grok care to make an estimate on when this will be released?
@cursor_ai Why from scratch? Is there value from “reinventing the wheel” something wrong with retraining from existing weights? Serious question not trying to be facetious