The DeepSeek-V4 preview, posted this fortnight to arXiv, is not a leaderboard story. It is a topology story. The headline number everyone will quote — 1.6 trillion parameters, 49 billion activated in the Pro model — is the least interesting line in the abstract. The load-bearing line is buried lower: in a one-million-token context, DeepSeek-V4-Pro reports needing only 27% of the single-token inference FLOPs and 10% of the KV cache of its own V3.2 predecessor.

Read that as an infrastructure engineer, not a benchmark-watcher. The KV cache — the running store of attention keys and values that every generated token has to consult — is the queue that actually breaks under load. It grows linearly with context length, it lives in the most expensive memory you own, and it is the reason a “million-token window” is usually a marketing ceiling nobody reaches in production. DeepSeek’s team attacks it structurally: a hybrid of Compressed Sparse Attention and Heavily Compressed Attention, sitting under Manifold-Constrained Hyper-Connections and trained with the Muon optimizer over 32T tokens. Strip the vocabulary and the move is simple — spend engineering to shrink the queue instead of buying more memory to hold it.

←TODAY: A million-token context stops being a demo and becomes a KV-cache budget you can actually afford. →3012: The office that reasons over its whole project at once — every Pset, every SIA clause, every RFI — not one PDF at a time. Fulcrum: Long-context intelligence was never gated by attention math; it was gated by the cost of remembering, and that cost just fell 10×.

Here is why this reaches your desk and not just a GPU cluster. The reason your BIM coordination assistant forgets the fire-rating decision from three prompts ago is the same reason DeepSeek-V3.2 was expensive at length: the model either drops old context or pays full freight to keep it. PAZ has covered the sharp end of this before — our archive note on Anthropic’s March 2026 cache-TTL regression, drawn from 119,866 logged API calls, showed a 12.5× cost multiplier the moment a cached context expired and had to be re-uploaded. That is the entire economics of long context in one number. Cheaper KV cache is the difference between an assistant that holds your project brief all day and one that reloads it, at write rate, every time you pause for coffee.

Atelier: For a Büro adopting AI coordination, the temptation is to chase the biggest advertised window. Resist it — the window you can afford to keep warm matters more than the one you can technically open. Monday move: instrument one real session. Log how many tokens your assistant actually re-reads per turn on a live IFC handover, and you will find the same silent queue DeepSeek just optimised. Measure before you scale the context.

The trade-off stays on the page: an open checkpoint at 1.6T parameters is not something a practice self-hosts, and “27% of the FLOPs” of a very large model is still a very large model. Efficiency here means the frontier labs can offer long context at a price that doesn’t collapse — not that the compute vanished. It moved, and someone still pays the cooling bill.

Hack: Compute the queue yourself before you trust anyone’s window. The KV cache is the dependency nobody draws on the architecture diagram, so size it directly — this is what “10% of the KV cache” is actually buying back in gigabytes.

layers, kv_heads, head_dim, tokens, dtype = 61, 128, 128, 1_000_000, 2
full_gb = 2 * layers * kv_heads * head_dim * tokens * dtype / 1e9
print(f"full KV cache: {full_gb:.0f} GB")
print(f"at DeepSeek-V4's ~10%: {full_gb * 0.10:.0f} GB")

Run it and the abstraction stops being abstract: the difference between those two numbers is the difference between a context that fits on your hardware and one that doesn’t. That is the whole exercise — find the third single point of failure you didn’t know you had, and here it is a memory queue with your project’s entire history in it. Draw your real dependency graph, put the KV cache on it, and decide what you can afford to remember before the vendor decides for you.

Source: arXiv