Summary
DeepSeek has released an updated DeepSeek V4.1 Flash checkpoint just 37 days after the original, with an average benchmark gain of 13.5%. It has also open-sourced the core inference engine behind its production serving stack. This video from bycloud explains why that release matters for anyone who runs or pays for large language model inference.
The centerpiece is DeepSeek’s claim that it earns about $38 per GPU-hour serving tokens, even at off-peak prices. The video compares that with the company’s February 2025 V3/R1 figures: roughly 1,814 H800 GPUs, 776 billion tokens in a day, and about $12.91 per GPU-hour. It then walks through where the gains come from, including the tradeoff between total throughput per GPU and per-user generation speed. One example is around 26,000 output tokens per second per GPU at 100 tokens per second per user, falling to about 16,000 at 250.
The video also covers wide expert parallelism, with ablations comparing 8-GPU and 64-GPU decoding setups, and explains why memory bandwidth limits mixture-of-experts decoding. It touches on KV cache reuse for agentic coding workloads with very long contexts. Viewers also get a brief look at the model’s reported performance on a multi-step offensive cybersecurity evaluation and what open weights mean for access to that capability.
📺 Source: bycloud · Published November 03, 2026
🏷️ Format: Deep Dive
