Why does DeepSeek activate only 3% of what it knows?

A Model That Knows More Than It Thinks With

DeepSeek V4.1-Flash holds 552 billion parameters — and here is the strange part. At any given moment, the model is actually using somewhere between 8 and 16 billion of them. The rest sit dark, waiting. You have built a library with half a million shelves, and you only ever reach for a handful of books at a time.

Those two numbers, 552 billion and 8 billion, are not a contradiction. They are the whole story. The 552 billion is what the model knows; the 8 billion is what it thinks with during a single operation, rising to 16 billion when it writes its answer back out. The gap between those figures is not waste. It is the architecture's central bet: that intelligence, at scale, does not require activating everything at once.

Four days after the model launched, DeepSeek made a quiet but telling decision. On September 14, they began routing all professional API traffic that had been headed for the older V4-Pro straight to V4.1-Flash. No fanfare, no migration guide plastered across the front page. The new model was simply good enough, and efficient enough, that the older one became redundant overnight.

The headline parameter count, 552 billion, is the kind of number that sounds like a boast. But the number that actually matters is the active one. A brain with 86 billion neurons does not fire all of them to read a sentence.

The art is in the routing. What DeepSeek built is a system that knows an enormous amount and chooses, precisely, what to wake up. That choice is what the rest of this story is about.

The Architecture of Selective Attention

Think of the model as a vast library with 552 billion books. Every query you send does not wake the librarian to search all of them. A fast routing mechanism reads your question at the door, selects the relevant shelves, and sends only that smaller crew to work.

For prefill — the processing of your input — that crew numbers 8 billion parameters. For decoding, generating the reply, it expands to 16 billion. The other 528 billion remain perfectly still, drawing no power, doing no computation.

This routing is the Mixture of Experts mechanism at the heart of DeepSeek-V4.1-Flash's Causal Encoder-Decoder architecture. Where a standard dense transformer applies every parameter to every token, the MoE router learns to assign tokens to specialist sub-networks, called experts, and ignores the rest. The Causal Encoder-Decoder design extends this further by separating the two jobs — understanding input and producing output — into distinct computational stages with different activation budgets. Separation is the key word. It lets the architecture be enormous in theory and minimal in practice.

The harder engineering problem is keeping such a sparse system stable. Activate only a fraction of a network on each forward pass and the topology can drift: gradients flow unevenly, some experts overtrain, others go cold. DeepSeek addresses this with Manifold-Constrained Hyper-Connections, or mHC, a set of learned constraints that preserve the geometric relationships between active and dormant regions even as only a thin slice fires at any given moment. The analogy, imperfect but useful: it is like a suspension bridge designed so that traffic on two lanes does not warp the cables serving the empty four.

Then there is the Internalizer. When the model encounters a long document, a hypernetwork generates a document-specific Low-Rank Adaptation, a LoRA adapter, tailored to that text in a single forward pass, without retraining the frozen 284-billion-parameter base. Devine et al. describe this as giving the model a temporary, disposable lens, ground fresh for each document, then discarded. It is adaptation without learning, context-sensitivity without weight updates. That distinction is quiet but significant.

How a Million Tokens Fit in a Slimmer Memory

Every conversation with a language model generates a hidden tax. As the model reads and processes text, it builds what engineers call a KV cache — a running record of everything it has seen, stored so it does not have to recompute from scratch at each step. In large models handling long documents, that cache grows enormous. It becomes the memory bill that limits what you can practically afford to ask.

DeepSeek's Compressed Sparse Attention 2 cuts that bill to one-eighth of what previous generations required. Not by throwing away information, but by finding smarter geometric shortcuts through the attention computation — keeping the meaning while shedding the bulk. Think of it as the difference between storing every photograph your camera ever took versus keeping a well-indexed album of the decisive frames. You lose very little; you gain back most of your shelf.

That freed capacity is then put to deliberate use. Engram Conditional Memory is a dedicated subsystem — 196 billion parameters whose sole job is retrieval across a context window of one million tokens. One million tokens is roughly 750,000 words, which is about seven full novels loaded simultaneously. For most models, a context that long is a theoretical boast. Here it is an engineering decision backed by a memory architecture specifically sized for the task.

For real enterprise work, this distinction is not trivial. A legal team reviewing a complex merger can load the entire document trail in one session. A software engineer can hand the model an entire large codebase. The model does not summarize or forget the beginning by the time it reaches the end. That is the practical test a benchmark cannot fully capture, and it is the one that changes how much a tool can actually be trusted.

When reasoning costs almost nothing, the question of who gets to use intelligence stops being a financial question.

Running on the Other Chip: DeepSeek, Huawei, and the Geography of Silicon

Picture a chip foundry in Shenzhen, the air thick with the smell of epoxy resin, and a very specific engineering frustration. When US export restrictions cut off Chinese laboratories from NVIDIA's most powerful accelerators, the problem was not simply one of computing muscle. It was cultural: the entire global ecosystem of AI training had grown inside CUDA, NVIDIA's proprietary software environment, the way a river carves its own bed. Switching chips meant not just swapping hardware but rebuilding the riverbed.

DeepSeek did exactly that. The V4 series was optimized to run on Huawei's Ascend 950PR chips using CANN Next, Huawei's heterogeneous computing framework and its answer to CUDA. That is a quiet sentence about an enormous engineering effort. CANN Next handles the gap between the way researchers write AI code and the way Huawei silicon actually processes it — a layer of translation that NVIDIA spent fifteen years making invisible and that DeepSeek's engineers had to make equally invisible, quickly.

The result is a frontier AI model that runs without a single NVIDIA GPU. Spend a moment with that fact. The assumption baked into most of the AI industry in 2024 was that the cutting edge of machine intelligence had a fixed address: Santa Clara, California. That address now has a forwarding label.

The investors noticed. In October 2026, DeepSeek closed a funding round of 80 billion yuan, roughly twelve billion US dollars, with Tencent and CATL among those placing bets. CATL makes batteries; Tencent runs half of China's internet. Neither invests carelessly. What they bought into was not just a capable model but a hardware-independent architecture — a proof of concept that the geography of AI silicon might, at last, be renegotiable.

Whether the Ascend 950PR can sustain massive-scale inference reliably over years remains, honestly, an open question.

What the Benchmarks Say About DeepSeek V4.1-Flash — and What They Don't

A score of 90.9 on GPQA Diamond sounds like a grade. It is more like a threshold. The GPQA Diamond benchmark consists of questions written by domain experts specifically to defeat people who only half-know the subject — graduate-level chemistry, biology, and physics problems that catch pattern-matchers and reward genuine reasoning. Humans with relevant PhDs score around 65 percent. A 90-something, from any model, is unusual enough to make researchers look twice.

DeepSeek-V4.1-Flash hit that number. Its predecessor, V4-Pro, scored 80.6 percent on SWE-bench Verified, a test that does not ask an AI to describe how to write software but actually makes it write software — find the bug, fix it, submit the patch, pass the tests. Eighty percent on that benchmark means the model is functioning as a working junior engineer on real open-source repositories, not performing a clever impersonation of one.

Then there is throughput: 333 to 427 tokens per second in practice. At the higher end, that is a full page of text every two seconds, fast enough that waiting stops feeling like waiting.

Now the honest part. Benchmarks are well-lit corridors, and models are trained to walk them well. An academic leaderboard says something true; it does not say everything. Comparisons to Claude 4.6 suggest DeepSeek-V4.1-Flash lands in similar territory on general tasks, but at a dramatically lower price point — a soft finding, not a controlled trial. What the independent research has confirmed more rigorously is the Pareto frontier result: in agentic optimization testing, V4.1-Flash and GPT-6 Astra currently occupy the edge where neither cost nor performance can improve without sacrificing the other. That is a precise claim about a narrow domain, which is exactly the kind of claim worth taking seriously.

The benchmarks are a map, not the territory. But this particular map has some interesting roads on it.

The Price of Thinking Is Approaching Zero

Run a million words through DeepSeek-V4.1-Flash and the bill comes to $0.22. That is less than a cup of bad coffee from a petrol station vending machine. Cache those same tokens for reuse and the price drops to a cent per million — a number so small it barely qualifies as a price at all.

To understand what this does to competitors, consider the margin structure of AI inference. OpenAI and Google have built businesses on the assumption that frontier reasoning commands a premium. DeepSeek's pricing removes that assumption surgically. Independent benchmarks place V4.1-Flash alongside GPT-6 Astra on the Pareto frontier of performance and cost — meaning you cannot get better value anywhere else without sacrificing one or the other.

The business model questions this raises are genuine and, honestly, still open. We do not yet know how sustainable $0.01 token pricing is, nor whether the $12 billion funding round carries invisible subsidies that distort the comparison. What we do know is simpler and more consequential: when reasoning costs almost nothing, the question of who gets to use intelligence stops being a financial question. DeepSeek V4.1-Flash has moved that frontier from theory into practice. That is where the real story starts.