Fine-tuning gets 75% cheaper while a 501B open model lands on serverless and the B200 price sheet goes public

Fine-tuning on a diet

Crusoe's announcement this week is a direct attack on the most annoying line item in post-training: memory. Its new method performs LoRA fine-tuning directly on 4-bit quantized base models using proprietary dequant-on-the-fly kernels, which removes the requirement to hold base weights in bf16 precision during training. The company says the result is up to 75% less GPU memory with no train-inference accuracy loss, and that the same approach lets large mixture-of-experts models train on up to 75% less hardware footprint. Every figure is Crusoe's own — the kernels are proprietary, so outsiders cannot verify the trick, only the results. m:…3026

The timing is not accidental. Crusoe notes that frontier models — it names DeepSeek-V4, Kimi-K3, and GLM-5 — are already being shipped heavily quantized from 16-bit to 4-bit weights. Fine-tuning infrastructure that assumes bf16 base weights is tuning for a world that is disappearing. The economic logic runs one way: as base models ship smaller, the adapters get cheaper to train, and the advantage moves to whoever has the data and the task, not whoever rents the most memory. m:…3026

The 501B model you can rent

Weights & Biases, meanwhile, made the case that the models themselves are getting bigger and more open at the same time. Reflection's Beam — a 501-billion-parameter mixture-of-experts model with 23 billion active parameters per token and an approximately one-million-token context window — is now available on CoreWeave Forge through Serverless Inference, released under an Apache 2.0 license. The combination is the point: half a trillion parameters, serverless, permissively licensed. m:…5450

CoreWeave did not just host the model; it validated the fine-tuning workflow, demonstrating supervised fine-tuning of Beam with Axolotl on its infrastructure — including an 84-million-token run on GPU kernel generation using PyTorch-to-Triton and CUDA examples that improved performance on specialized datasets. A day-zero fine-tuning recipe for a 501B open model, on rented GPUs, is the kind of thing that was a research project two years ago and is now a launch checklist item. m:…5450

The price sheet goes public

And then there is the hardware itself, priced like a commodity. fal's 2026 guide to the NVIDIA B200 — the first Blackwell-generation datacenter GPU, with 192 GB of HBM3e memory and 8 TB/s of bandwidth — lists $7.99 per hour at list rates and as low as $5.49 per hour for committed capacity, billed per second while a runner is active. Publishing the price sheet matters more than the prices: memory-bandwidth-heavy inference for large models, video generation, and FP4-quantized workloads now has a public unit economics reference. m:…5799

Why it matters

Stack the three and the direction is unmistakable: adapting big models is becoming a normal line item instead of a capital project. Quantized fine-tuning cuts the memory bill by three quarters, open 500B-class models arrive with day-one serverless recipes, and the flagship GPU's hourly rate is now public information you can plan around. The second-order effect is who gets to play: when training an adapter costs memory you can rent by the second, the scarce resource is the dataset and the eval, not the cluster. Budget for data and measurement before you budget for GPUs. m:…3026 m:…5450 m:…5799