how many tokens can a b300 with deepseek v4 flash prcoess a day with very good topmzaitons et

A B300 running DeepSeek V4 Flash can plausibly process on the order of tens to low hundreds of millions of tokens per day , but the exact number depends heavily on whether you mean input, output, or total tokens, plus batching and serving setup. One public estimate discussed for B300-class serving is about 800 generated tokens per second in a multi-user setup, which works out to roughly 69 million generated tokens per day at full saturation.

Practical estimate

If you take that 800 tokens/second figure as a rough throughput anchor:

• 800×60×60×24≈69.1800\times 60\times 60\times 24\approx 69.1800×60×60×24≈69.1 million generated tokens/day.
• That is for generated/output tokens , not counting all the extra input/context tokens that can also be processed.
• Real-world daily throughput can be higher or lower depending on prompt length, sequence length, concurrency, KV cache behavior, and whether the system is compute- or memory-bound.

Why the number varies

DeepSeek V4 Flash is described as a 284B total / 13B activated MoE model, which is designed for efficiency, and it supports a 1M token context with up to 384k output tokens in the published description. That means a B300 can be very fast for short-to-medium requests, but long-context workloads can reduce effective daily token volume because each request consumes more memory and compute per token.

Rule-of-thumb ranges

A useful rough framing is:

• Light workloads: well above 70M tokens/day equivalent.
• Mixed workloads: around 30M–70M tokens/day.
• Heavy long-context workloads: can drop below that significantly, sometimes by a lot.

Bottom line

For a B300 with DeepSeek V4 Flash and good optimizations, a defensible quick estimate is about 70 million generated tokens per day at sustained saturation , with a broader practical range of tens of millions to possibly over 100 million total tokens/day depending on workload shape.