Leadership1 publisher3 min readPublished
DeepSeek's V4 preview cuts million-token KV cache to a tenth of V3.2's
The efficiency claims are self-reported and measured against DeepSeek's own prior model, but the weights are MIT-licensed and already pulled 1.78 million times, which is what turns a ratio into a number a buyer can carry into a renewal.
The Board Room · Leadership desk
What happened
- DeepSeek released a preview of its V4 series as two mixture-of-experts models, Pro at 1.6 trillion parameters with 49 billion active and Flash at 284 billion with 13 billion active, both taking a million-token context.
- Both models were pre-trained on more than 32 trillion tokens, then post-trained by cultivating domain experts with SFT and GRPO reinforcement learning before consolidating them through on-policy distillation.
- The repository and the weights are released under the MIT License, which permits commercial self-hosting without a negotiated agreement.
- The Flash page carries leaderboard entries including GPQA Diamond at 88.1, SWE-bench Verified at 79 and Terminal-Bench 2.0 at 56.9.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- constraint FLOPs and KV cache ratios are not prices, and with no per-token rate published a buyer can quote the 10% figure but cannot compute a bill from it.
- capability If the cache figure holds on real hardware, the same memory footprint carries roughly ten times the concurrent million-token sessions, which changes what an in-house deployment can serve rather than only what it costs.
- decision A Flash tier that DeepSeek says matches Pro's reasoning given more thinking time gives a buyer a smaller self-hosted fallback to price against a long-context contract, and the incumbent has to answer it.
- contradiction The release makes its strongest claim, best open-source model available, on the evidence it documents least, since the listed scores are not tied to a named variant or reasoning mode.
KV cache is the part of long-context serving that ends up on an invoice. It is memory held per session for the life of the session, so it sets how many million-token conversations a given box can hold at once, while FLOPs per token mostly govern latency and utilisation. DeepSeek reports both falling at a 1M-token context, to 27% of DeepSeek-V3.2's single-token inference FLOPs and 10% of its KV cache, credited to a hybrid of Compressed Sparse Attention and Heavily Compressed Attention [3][2]. Of the two, the memory number does more work: one tenth of the cache is roughly ten times as many concurrent 1M-token sessions per unit of memory before some other bottleneck binds [4], against a 3.7x reduction in per-token compute [5].
Nothing here says scale got cheaper to build. V4-Pro carries 1.6 trillion parameters and both models were pre-trained on more than 32 trillion tokens [1][4]. The saving is per token at long context, and the comparison point is DeepSeek's own V3.2 rather than any incumbent's served product [3], which means the ratio travels only as far as a reader is willing to treat V3.2 as a stand-in for what they are paying for now.
Sparsity is where the serving economics sit. Pro activates 49 billion of its 1.6 trillion parameters per token, about 3.1% [1]; Flash activates 13 billion of 284 billion, about 4.6% [2], roughly 3.8 times fewer active parameters than Pro [3]. Expert weights are stored in FP4 with most other parameters in FP8 [8], and DeepSeek says the Flash maximum-reasoning mode reaches comparable reasoning performance to Pro when given a larger thinking budget, trailing on pure knowledge and the most complex agentic workflows [7]. That is a trade of thinking time for hardware, and it is the tier most in-house deployments would actually be sizing.
A skeptic would say a model card is a marketing document, and this one is a preview at that [1]. The answer is that the weights and repository are MIT-licensed and the Flash page already logs 1,779,052 downloads in the last month across 100 Spaces [9][10], so anyone with the hardware can test the claim. What the record does not yet carry is the 27% and 10% reproduced on named hardware at named batch sizes, or a score tied to a named variant: the page lists GPQA Diamond at 88.1 and SWE-bench Verified at 79 [12] without the text saying which model and which of the three reasoning modes produced them [13][6].
For this quarter the usable fact is optionality rather than a discount. The release documents licensing, sampling settings and a recommended 384K minimum context window for the maximum-reasoning mode [11], and states no per-token price [14]; it also ships without a Jinja chat template, offering encoding scripts instead, which is real integration work for anyone swapping it in [15]. The durable part, per the paper title DeepSeek attached to the release, is the claim that million-token context is an efficiency problem [16]. A vendor asked to defend long-context pricing now has to answer a published architecture, not a hunch.
What to watch
- An independent reproduction of the 1M-token FLOPs and KV cache figures on named hardware, reported as tokens per second and sessions per GPU.
- A per-token price for V4-Pro or V4-Flash, from DeepSeek or a third-party host, that can be set beside incumbent long-context rates.
- An outside evaluator publishing GPQA or SWE-bench results attributed to a specific V4 variant and reasoning effort mode.