DeepSeek's V4.1-Flash trims KV cache to 890 bytes per token
reported by 2 outlets· 2 articles · confidence: high · first seen 2026-09-10 07:31 UTC
What this means for you
If long-context agent traffic is your cost problem, this is testable now: the weights are public and the claim is a quarter of V4-Flash's per-token cache. No hosted pricing, no third-party numbers and no benchmark date, so the saving is asserted, not measured.
DeepSeek released V4.1-Flash, a multimodal mixture-of-experts model (only part of the network runs per request) with a 552B-parameter backbone and a one-million-token context window. It targets the memory long agent sessions consume: a KV cache — the per-token state a model keeps so it need not reread earlier text — of 890 bytes per token, about a quarter of V4-Flash's, and roughly an eighth on disk. It activates 8B parameters per token during prefill and 16B during decode, and was pretrained on 45T tokens. Weights are on Hugging Face. The paper claims better performance than V4-Flash without giving scores, an evaluation date or a harness.
Key facts
- ·DeepSeek-V4.1-Flash is a multimodal mixture-of-experts model with a 552B-parameter backbone and support for contexts up to one million tokens. source
- ·Under its Causal Encoder-Decoder architecture it activates 16B parameters per token during decode and 8B during prefill. source
- ·Global KV cache footprint is 890 bytes per token, roughly a quarter of DeepSeek-V4-Flash's; the persistent footprint on SSD or host memory is about an eighth. source
- ·The model was pretrained on a 45T-token multimodal corpus, with checkpoints published at huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash. source
- ·The paper claims substantially better performance than DeepSeek-V4-Flash but reports no scores, evaluation dates or harness names. source
- ·MarkTechPost reports a further 196B Engram parameters alongside the 552B backbone; the arXiv paper does not mention them. source
What the sources say
- Hugging Face Daily Papers (research) — Primary paper: architecture, the 890-byte-per-token cache, 45T-token pretraining, and where to get the weights.
- MarkTechPost (press) — Press write-up adding a 196B Engram parameter count and restating the context window and cache design.
Sources
The original reporting. Follow these — they did the work.
- MarkTechPostDeepSeek AI Released DeepSeek-V4.1-Flash with 1M Context, FP4 KV Cache, and Cross-Layer Attention Reuse2026-09-10
- Hugging Face Daily PapersDeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression2026-09-16