There are two ways to build a more powerful AI right now. The first is the one everyone uses. You buy more GPUs, build bigger data centers, and burn hundreds of millions of dollars on electricity and raw compute. OpenAI, Google, Anthropic, and Elon Musk have spent three years striving to accomplish this. It’s what we in France and Europe keep failing to replicate.
More power, more money, more infrastructure — brute force all the way down.
And then there’s the second way. The one DeepSeek just demonstrated.
DeepSeek is a Chinese AI lab that doesn’t have access to the best Nvidia chips — U.S. export sanctions block them — and doesn’t have OpenAI’s budget. What they do have, apparently, is a willingness to ask questions no one else was asking. On September 10, 2026, they published DeepSeek V4.1 Flash: an open-weights model, MIT license, available to anyone on Hugging Face. Within hours, the community had already published modified versions.
A speed record for that kind of adoption in 2026.
The model competes with GPT-5.6 Sol and Claude Fable 5 on most benchmarks. Its inference cost is 40 times lower. One key architectural decision involved deleting memory that every other lab in the world was carefully saving.
I want to walk you through what they actually did, because the logic is genuinely elegant, and because understanding it changes how you think about where local AI is heading.
The Problem Every Lab Has Been Ignoring
To understand what DeepSeek solved, you first need to understand the problem they were solving. It has a name: the KV cache.
When you send a prompt to an AI model, two things happen. First, the model reads everything you gave it — your text, your documents, your context — and at each layer of its neural network, it takes notes. These notes are pairs of numerical values called keys and values. Together, they form what’s called the KV cache: a vault of everything the model knows about your prompt. Then, in the second phase, the model writes its response, token by token. And to pick each word, it has to re-read those notes from the first phase. The KV cache saves it from recalculating everything from scratch at every single token.
For a short prompt, no problem. The notes fit easily in the GPU’s fast memory — the VRAM soldered directly next to the processor, where access is essentially instant.
But the industry hasn’t been working with short prompts since 2022. Today we want autonomous AI agents that run for hours, processing hundreds of pages of documentation, entire codebases, and conversations spanning dozens of turns. The notes don’t fit anymore. The KV cache overflows out of VRAM and spills onto SSDs or regular RAM — wherever there’s space. And every time the model needs a note that’s been pushed out there, it has to ferry that data back across the entire motherboard to the GPU. The processor sits idle, waiting for the transfer. That waiting is why models slow down at long contexts. It’s also why high-end graphics cards with massive VRAM have become the hottest commodity in tech — more memory means more notes staying close to the processor.
The industry has been buying more. Bigger GPUs, faster interconnects, data centers with electricity bills that make your eyes water. Elon Musk is building the largest building on the planet and filling it with compute cards.
Brute force.
DeepSeek had to find another way. And their first idea was almost offensively simple.
The Cut That Halved the Reading Cost
A standard 40-layer model reads your entire context 40 times — once per layer. DeepSeek asked an obvious question that apparently no one had asked before: why does every layer need to read?
Their answer: it doesn’t. So they cut the model in half. The first 20 layers — the encoder half — read the context and build the KV cache. The last 20 layers — the decoder half — read nothing. They pull directly from the encoder’s notes and focus exclusively on writing the response, maintaining coherence by keeping a very precise eye on the last few words they’ve just written.
This architecture is called a Causal Encoder-Decoder, or CED. And it shouldn’t work as well as it does. A model where half the layers are blind to the original input is, on paper, a dumber model. But the decoder isn’t blind — it has full access to everything the encoder captured, plus tight attention to what it’s currently generating. The result: the reading phase costs exactly half what it used to. Not by buying faster hardware. By stopping 20 redundant read operations.
Clever. But it doesn’t solve the entire problem. The KV cache itself is still enormous the moment context grows. You’ve halved the compute cost of reading, but the notes still take the same amount of space.
That’s where the second discovery comes in.
From 40 Note Sets to 4
Normally, each of the 40 layers in a model creates its own set of notes. Forty layers, forty KV caches, enormous redundancy. DeepSeek’s second decision: only 4 of the 40 layers are allowed to create their own notes. The other 36 have to work with what already exists.
Going from 40 KV caches to 4 is not a small optimization. DeepSeek published the numbers: global KV cache memory drops from 3,514 bytes per token to 890 bytes per token — roughly a 4× reduction. What once demanded an entire server rack’s setup can now fit on hardware designed for a single developer or a small business. I’ve shown you how to set this kind of thing up. You can see exactly where this is going.
“Any developer, any lab, any country can download this model and build on top of it.” — DeepSeek release announcement, September 10, 2026
However, the paper had yet to discuss one more kind of memory. And the solution to that one is the most counterintuitive decision in the whole paper.
The Decision That Looked Insane
The decoder half of the model — those 20 layers that don’t read the original context — keeps a very precise focus on the last 128 tokens it’s generated. Think of it as short-term memory: what was just said, to maintain sentence-to-sentence coherence. In a multi-turn conversation, after each response, the model had to save that local memory to disk to carry it into the next turn. Standard practice everywhere.
That constant saving was saturating storage. So DeepSeek did something that sounds completely backward: they deleted it.
Once the model completes generating a response, the local memory is cleared.
Gone.
The model stays consistent since the GPU recomputes those 128 tokens from scratch at the start of every new turn. This mechanism is called DSpark. For a modern GPU, recalculating 128 tokens takes microseconds. It’s negligible. Fetching those same tokens from a disk, transferring them across a motherboard, loading them into the processor — that takes orders of magnitude longer.
