Magic2026-09-09 08:40:52Magic says $500,000 in GB200 compute can approach DeepSeek V4 Pro Base pretrainingMagic, an AI coding company, said a new pretraining setup can achieve results close to DeepSeek V4 Pro Base using about $500,000 worth of NVIDIA GB200 compute, with total compute demand at roughly 1/50 of the comparison benchmark. The company said the comparison targets a base model that has only gone through pretraining and has not yet received reinforcement learning or other post-training work. That means the result is not presented as evidence that its chat, coding, or agent performance has caught up with the finished DeepSeek V4 Pro product. Magic said it later scaled the training run by 10x, or about $4 million in GB200 compute. In the same internal evaluation, that larger version outperformed the latest public base models from DeepSeek, Kimi, and NVIDIA that it tested. The company also estimated that reaching the same level at DeepSeek V4 Pro’s training efficiency would cost more than $100 million in compute. Magic did not release the full training recipe. It only said the efficiency gains came from dozens of changes across model architecture, optimizer design, training objectives, and data processing. Model weights have not been published either, leaving the claimed 50x efficiency gain without independent external reproduction for now.860
Stanford2026-08-24 13:07:10Stanford professor Percy Liang opens Marin 535B model training to the publicStanford professor and Simile AI founder Percy Liang has kicked off training for Marin 535B-A23B and is making the process public in real time. The model is described as having 535 billion total parameters and 23 billion active parameters, trained on 18.75 trillion tokens across 11 GB200 NVL72 systems, or about 792 GB200 GPUs. Liang said the run is expected to last roughly three months and consume about 2.7e24 FLOPs, followed by a post-training phase. Before starting the main run, the team completed a four-stage scaling ladder ranging from 1.6B-A61M with 48 billion tokens to 27.7B-A1.2B with 926 billion tokens. Liang said those smaller runs were used to surface and debug issues early and to forecast the behavior of the larger “hero run.” The project has drawn attention because it exposes details that are usually hidden in frontier model development, including data mix, processing methods, training configuration, code, experiment design, live loss curves, and model-state tracking. The team has also published links to a data overview page, a Weights & Biases dashboard, and a GitHub issue page where the work can be followed in detail.1860
