# DeepSeek's V4.1-Flash paper cuts the global KV cache to 890 bytes per token for a 552-billion-parameter model

Cross-layer cache reuse and 4-bit (FP4) storage of the key-value (KV) cache cut the cache held in high-bandwidth memory (HBM) to about a quarter of V4-Flash's, the weights ship under an MIT license, and the result lowers the price floor that closed APIs charging $4 and $1.25 per million input tokens must compete against.

- Published: 2026-11-01T06:20:01.798Z
- Canonical: https://polylog.news/ai/2026-11-01/deepseek-s-v4-1-flash-paper-cuts-the-global-kv-cache-to-890
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [arXiv](https://arxiv.org/abs/2609.19969), [Hugging Face Papers](https://huggingface.co/papers/2609.19969), [LLM Gateway](https://llmgateway.io/timeline), [Meta AI](https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/), [AI Weekly](https://aiweekly.co/alerts/meta-prices-muse-spark-11-api-at-125425-per-m-tokens), [Anthropic News](https://www.anthropic.com/claude-opus-5-5), [LLM Stats](https://llm-stats.com/models/claude-opus-5-5)

DeepSeek's technical report on V4.1-Flash describes a multimodal mixture-of-experts model with 552 billion backbone parameters and a one-million-token context. It activates 8 billion parameters while processing the initial prompt (prefill)…

This story is for subscribers. Read it in full at https://polylog.news/ai/2026-11-01/deepseek-s-v4-1-flash-paper-cuts-the-global-kv-cache-to-890 (subscription information: https://polylog.news/pricing).