LoRA freezes the base weights. So why store them in bf16?
We keep them in mxfp4 in HBM and dequantize inside the kernel. The optimizer only touches the LoRA adapters.
Up to 75% less base-weight memory, with no accuracy loss in our internal evals.
➕ The math:
Last day at @CloudNativeFdn's #KubeCon# + #CloudNativeCon#, let's go!
🎥 A look from the convention center floor. Stop by the Crusoe booth to meet the team, test your technical knowledge, and win prizes!
Crusoe is on the inaugural 2026 AI 100 from @NewcomerMedia and @BessemerVP.
Congrats to our fellow honorees building the future of AI, including those we're fortunate to call customers. We're proud to power their work.
See the full list: 🔗
In Salt Lake City for @KubeCon + @CloudNativeCon? Join Crusoe TONIGHT for Containers & Cocktails!
📅 Nov 11 | 🕕 6pm MT
Enjoy curated cocktails, great food, and conversation with industry experts.
Space is limited! Register:
#CloudNativeCon# #KubeCon#
From service to servers. 🇺🇸 To every veteran — thank you for your service.
Your next mission. Build the future together. We’re hiring veterans.
Apply:
Reach us: veterans@crusoe.ai
Know a veteran who’d thrive here? Send them this post.
#VeteransDay# #HireVeterans#
🎥 Behind the scenes building our @KubeCon + @CloudNativeCon booth!
We’re in SLC through Thursday. Stop by to book a 1:1 with our technical team.
📅 Nov 12, 12:40 PM: @CrusoeDev's Connor Guerrero presents "When Should Your Cluster Fix a Failing GPU?"