DeepSeek formally released V4.1-Flash today. Open weights under MIT licence on Hugging Face, with vLLM and SGLang support.
The specs: 552 billion parameters, multimodal mixture-of-experts, including 196 billion Engram parameters. Active parameters are 8 billion during prefill and 16 billion during decode. Context window: one million tokens.
The number that matters
890 bytes of KV cache per token. That’s about 25% of what V4-Flash needed, and a fraction of V1 – DeepSeek puts the factor at 437.
Three things combine to get there. The main KV cache sits in FP4 (E2M1 with E4M3 scales per 16 channels), halving storage versus FP8. Sliding-window attention is no longer held in full but rebuilt by replaying the last 128 prompt tokens. And across layers the model runs in three modes: one computes its own KV and picks new top-512 indices, one reuses KV but rescores it with its own queries, one skips indexing entirely. DeepSeek calls it CSA2.
On benchmarks, the headline results include 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1.
Pricing, and the end of V4 Pro
From September 14 at noon Beijing time, requests to V4 Pro get routed to V4.1-Flash and billed at Flash rates. DeepSeek’s reasoning: Flash beat its predecessor on performance, cost, speed and total completion time in both internal and external testing.
Off-peak pricing: 0.02 yuan per million tokens on cache hits, 1 yuan without, 4 yuan for output. Peak hours cost double.
Why memory beats the benchmark here
If you run agents, you know the shape of this problem. KV cache grows with every step, and at some point the constraint isn’t the model, it’s the memory on the card. On long runs with lots of tool calls, that’s the wall you hit first – not model quality.
A quarter of the memory means roughly four times as many parallel sessions on the same hardware, or considerably longer runs before anything falls over. If you self-host, that’s the more useful headline than any point on Terminal-Bench.
And yes, it fits the pattern you’ve been able to watch for months now. Chinese labs optimise against memory and cost because they can’t just throw compute at the problem. That constraint is currently producing the more interesting architecture ideas.
Whether the numbers hold up, we’ll know in a few days once people measure independently. The weights are out there.
Sources: TechNode: DeepSeek formally launches V4.1 Flash, routes V4 Pro requests to Flash · MarkTechPost: DeepSeek AI Released DeepSeek-V4.1-Flash with 1M Context, FP4 KV Cache, and Cross-Layer Attention Reuse