Compare Open Source VLMs
DeepSeek V4.1 Flash is a 552B-parameter MoE model, with substantially fewer active parameters per token. Metric Size Backbone parameters 552B Engram memory parameters 196B Total checkpoint parameters, including auxiliary components ~763B Active parameters during input processing 8B Active parameters during generation 16B Published checkpoint size ~510 GB GPU memory for deployment ~614 GB Important distinction: It has the compute characteristics of an 8–16B active-parameter model, but the memory requirements of a much larger model. For your robotics RL experiments, that's a major limitation. Full-model fine-tuning would be expensive, despite its efficient inference. Open weights on Hugging Face (MIT license). Yes. For real-time robotics, I'd target a 1–3B VLM/VLA student, not 27B. You can go down to 0.2–0.8B for more specialized control policies. The central idea is to use Astra-scale intelligence during training, but a much smaller, faster model for deployment. A 27B student makes sens