DeepSeek Ships V4.1-Flash, Compressing KV Cache to 890 Bytes per Token

DeepSeek released V4.1-Flash, a multimodal API model emphasizing faster inference, higher throughput and lower pricing. Its headline feature is long-context efficiency: architectural changes compress KV cache memory to just 890 bytes per token, enabling larger context windows and substantially lowering inference memory and cost, drawing rapid community teardowns and benchmarks.

Confirmed

The Prompt Engineering channel dissected the architecture, identifying CED separation, CSA2 layer sharing and 4-bit quantized replay as the core compression techniques, and judged it possibly the best local vision model today.
TensorSharp, an open-source inference engine, added a dedicated execution path for V4.1 Flash; author fuzhongkai benchmarked on 8×A40 (layer split, F16 KV cache, 65K context): Q2K prefill 533–539 tok/s, single-stream decode 40 tok/s; Q4KM prefill 45x tok/s (truncated in source).
Reddit user @justlikemedics analyzed that the method significantly reduces memory requirements and cost for inference serving.
Blogger @thione noted V4.1-Flash continues DeepSeek's value-for-money路线, with pricing and benchmarks on the official release page.

Unconfirmed

Exact pricing and official benchmark figures are not in the source posts; the Q4KM prefill number is truncated.

Why it matters

KV cache is the biggest memory bottleneck for long-context inference; 890 bytes per token means longer contexts and higher concurrency on the same hardware, directly cutting deployment costs.
Fast adoption by open-source tooling like TensorSharp shows a usable self-hosted ecosystem already exists.

2026-09-13 ~ 2026-09-15 · 6 related posts

Primary sources

DeepSeek V4.1 Flash: KV-cache shrunk to 890 bytes per token — Prompt Engineering · 2026-09-14 21:04 (UTC+08:00)
DeepSeek V4.1 Flash Hits 40 tok/s on 8x A40 with Open-Source TensorSharp Engine — fuzhongkai · 2026-09-13 01:57 (UTC+08:00)
DeepSeek releases V4.1-Flash, a multimodal API model with faster inference and lower prices — thione · 2026-09-15 00:15 (UTC+08:00)

[source] DeepSeek V4.1 Flash Hits 40 tok/s on 8x A40 with Open-Source TensorSharp Engine — fuzhongkai · 2026-09-13
TensorSharp Hits 500+ tok/s Prefill for DeepSeek V4.1 Flash on 8×A40 GPUs — fuzhongkai · 2026-09-13
DeepSeek V4.1 Flash cuts KV-cache to 890 bytes per token for cheap long context — Prompt Engineering · 2026-09-13
DeepSeek's V4.1-Flash KV cache compression could undercut OpenAI and Anthropic's compute moat — justlikemedics · 2026-09-13
[source] DeepSeek V4.1 Flash: KV-cache shrunk to 890 bytes per token — Prompt Engineering · 2026-09-14
[source] DeepSeek releases V4.1-Flash, a multimodal API model with faster inference and lower prices — thione · 2026-09-15