Need two to run the glm/DeepSeek class of models. Even with a pro 6000 which is more expensive than the sparks the best you can run is qwen27b which is pretty mid. They also get faster the more you add to the cluster with tensor parallelism