DeepSeek V4.1-Flash Cuts KV Cache to a Quarter: Is It Worth Switching?

2026-09-12 · 3 min read

deepseekmodel-releaseagentscomparison

What's new

DeepSeek released V4.1-Flash on September 10, 2026, open-sourced under the MIT licence on Hugging Face, with API pricing unchanged from V4-Flash. The selling point this time isn't benchmark scores. It's memory.

V4.1-Flash moves to a split encoder/decoder design: about 8B parameters are active per token while reading input and 16B while generating, out of 552B total, so input-side compute is cut roughly in half. The bigger change is the KV cache. The resident (in-VRAM) portion shrinks to about 25% of V4-Flash's; the portion that can be offloaded to slower storage shrinks to about 12.5%; and cache precision drops from FP8 to FP4, nearly halving footprint again. DeepSeek's own framing: per-token global KV cache is 437x smaller than V1. The context window is 1M tokens, and the training set is 45 trillion tokens including images.

The benchmarks show a trade-off. DeepSWE v1.1: 74.2%, edging past Anthropic's Opus 5 and OpenAI's GPT-5.6 Sol. But ProgramBench lags clearly, and complex scientific reasoning and image understanding are weaker. Reasoning depth is adjustable; the maximum setting emits roughly 2.5x more tokens.

Why the memory number matters

Cutting KV cache to a quarter is invisible to someone chatting in a browser. For AI agents it's a dividing line.

An agent's typical shape is a long task: dozens or hundreds of tool calls, context climbing into the hundreds of thousands of tokens. The cost bottleneck on that workload has never been compute — it's VRAM. When the cache doesn't fit, you either recompute or truncate, and truncation is how an agent "forgets" what it was doing. A 1M window with a quarter of the cache footprint means one GPU can run more concurrent agent sessions, and long tasks no longer need aggressive context compression to survive.

That's also why DeepSWE (real repository-level engineering tasks) is where V4.1-Flash overtakes Opus 5 while the more static ProgramBench falls behind: it's optimised for running long, not for answering cleverly.

It's already live on AIWITH.CHAT

We wired it up on release day. The DeepSeek robot on AIWITH.CHAT now defaults to V4.1 Flash; pick it from the version dropdown, with V4 Pro as the other option. No API key, no separate DeepSeek subscription — it shares one $9.9/month account with GPT-6 Astra, Claude Fable 5.1, Gemini 3.1 Pro and Kimi K3.

If you want to see how V4.1-Flash handles long documents, long conversations and multi-step agent-style tasks, open it and go; when you need heavier reasoning, switch to V4 Pro or GPT-6 in the same conversation.

See plans →

Is it worth switching?

Two kinds of reader, two straight answers:

You build agents, run long tasks, or deploy models yourself — yes, and it's high priority. Same price, same generation, a quarter of the memory. That's a pure cost reduction with nothing to weigh against it, and the MIT licence means you can swap the weights in without waiting for anyone.
You use AI as an everyday assistant — you don't need to do anything; the site is already on V4.1 Flash. The savings go to whoever pays for the GPUs, not to you, and the model is still weaker than its peers on complex scientific reasoning and image understanding. It's fine for Q&A, writing and long-document work; for heavy reasoning or vision, switching to GPT-6 or Claude in the same subscription is the better use of your beans.

In one line: V4.1-Flash is a release that lowers infrastructure cost, not one that raises end-user capability. Whether to switch depends on whether you're the one paying for VRAM or the one paying for a subscription.