A few days ago, we announced ~1,740 TPS decode for DeepSeek V4.1 Flash on 4xB300.
AI-generated summary Baseten reports a decode performance of approximately 1,740 tokens per second for DeepSeek V4.1 Flash running on 4xB300. Baseten states that the model runs up to 1.7x faster on Vera Rubin and notes that it is currently working on open-sourcing the implementation.