The exact KV cache usage of DeepSeek V4

Figure 1 of DSV4 paper seems to imply that DSV3.2 uses ~50GB at 1m context and DSV4 uses ~5GB: ***Numbers updated with the KV cache breakdown from vllm*** From my own calculations, the correct FP16 KV cache at 1m context should be: Model Params 128k 160k 1m KV% V3/3.1 671B 8.58GiB 10.72GiB 68.63GiB 5.11% V3.2 671B 10.48GiB 13.11GiB 83.88GiB 6.25% V4 Flash 284B 0.84GiB 1.05GiB 6.72GiB 1.18% V4 Pro 1600B 1.20GiB 1.50GiB 9.62GiB 0.3% So while KV cache saving is not 9.5x but 7.879x. It is still very

More from Reddit

Luce DFlash: Qwen3.6-27B at up to 2x throughput on a single RTX 3090

Reddit

GBNF grammar tweak for faster Qwen3.6 35B-A3B and Qwen3.6 27B

Reddit

Tissue regeneration moves closer to reality with lab-grown bone, muscle, and cartilage

Reddit

AI bill would crack down on deepfake distribution and protect whistleblowers

Reddit

India’s power demand hits record as heat wave batters nation

Reddit

Agents for end-to-end document redaction and review tasks (OCR and PII identification - Qwen 3.6 vs closed-source comparison)

Reddit