DeepSeek V4.1 Flash's Insane Architecture: Shared Memory That Shrinks KV Cache 437x

DeepSeek V4.1 Flash sits at the center of Two Minute Papers' 'Insane New Architecture' episode, which opens with sheer speed and two bold claims: on some benchmarks Flash can edge past Claude Opus-class models and Kimi K3, and it reliably beats DeepSeek V4 itself. The host frames cost to make the point tangible — running V4 Pro locally takes roughly $300k in hardware, while Flash does the same work for about a quarter. The curve points toward a pocketable agent within a year, offered as cautious optimism rather than hype.

What got faster, what got smaller: the anatomy of 437x

The most striking idea is that the model is huge and small at once. Huge because the backbone is 552 billion parameters with up to one million tokens of context — the video flags 'more than 500 billion.' Small because the KV cache that holds the context collapses: 437x smaller than V1 from three years ago and 4x smaller than V4 Flash from just months ago. The paper abstract in arXiv 2609.19969 pins the number: global footprint down to 890 bytes per token, roughly one quarter of DeepSeek-V4-Flash in HBM. With SWA Bounded Replay, the persistent footprint on SSD or host memory falls to about one eighth of V4-Flash. For long-horizon, input-heavy agent workloads, that directly cuts the bottleneck of compute, storage and bandwidth.

How is the shrink done? The video names it: Compressed Sparse Attention 2, CSA2. In a classic Transformer each layer keeps its own KV memory; DeepSeek V4.1 Flash shares it across layers. Not everyone has to remember everything — there is one shared memory. The difficulty is that different layers need different views of the same history. The fix is a 40-layer Causal Encoder-Decoder, CED: a 20-layer causal encoder projects a global memory from its final hidden states, and a 20-layer decoder reads from that global memory instead of deriving KV from its own states. Add FP4 KV caching and both the HBM-resident global cache and the SSD-resident persistent cache drop sharply.

Architecture numbers: 8B prefill, 16B decode, 45T tokens

The asymmetry matters for agent economics. The model activates 16B parameters per token during decode but only 8B during prefill, which halves prefill cost for input-heavy jobs. DeepSeek reports streamlining the V4 architecture, adding efficient extensions, and pretraining on a 45T-token multimodal corpus followed by broad post-training for text and multimodal agent scenarios. The 10 September 2026 announcement on deepseek.com positions it as the smallest member of the new family with native visual understanding, promising faster inference and higher throughput. Checkpoints are open on Hugging Face.

The video does not stop at architecture; it shows two concrete capability probes. First, native visual understanding — feed an iconic game menu image and have the model write a game that reproduces it. Second, physics reproduction — on the honey-coiling paper, GPT-6 Astra nails the result as 'absolutely stunning,' Opus 5.1 gets the physics right with slightly weaker rendering, and Flash is framed as a few papers away from matching that level. The host stresses showing the truth beyond headlines and notes, with a wry aside, that YouTube has stopped recommending his physics sim papers.

The catch and the bill: a model that thinks a lot and burns tokens

The catch at the end rebalances the story. V4.1 Flash likes to think and burns a lot of tokens. It is not expensive per token, but the sheer volume is high. With hardware you can run it for free; without, Lambda's GPU cloud is presented as the practical bridge — the host notes he uses Lambda to reproduce papers in minutes and to test ideas quickly. Every new DeepSeek paper, the argument goes, lowers inference cost for everyone, which directly helps doctors and scientists. Open weights plus an open paper keep the 'open science for the win' thread intact.

Stepping back, the insanity of V4.1 Flash is not raw muscle but memory economics. Cross-layer sharing, encoder-to-decoder projection and FP4 compression together shrink a one-million-token context to an 890-byte per-token trail while pushing performance above the baseline. Cutting a $300k local barrier to a quarter and halving prefill activation turns the 'in your pocket in a year' line from speculation into an engineering roadmap. By telling the shrink on two scales at once — 437x versus V1 and 4x versus V4 Flash — the video gives a compact frame for anyone tracking architecture.