DeepSeek's V4.1 Flash Cuts Memory Cost Fourfold, Which Changes the Self-Hosting Math
DeepSeek released a model this week that its own maker is using to shut down a bigger, more expensive model it sells. V4.1 Flash, announced September 10, carries a 552 billion parameter backbone but activates only 8 billion parameters to process an incoming prompt and 16 billion to generate a response. According to DeepSeek's technical report, the model outperforms V4 Pro, a 1.6 trillion parameter system with 49 billion active parameters, across most of the coding, cybersecurity, and autonomous agent benchmarks the company tested. The practical consequence arrived within the same announcement: starting September 14 at noon Beijing time, every API request sent to V4 Pro will be rerouted to V4.1 Flash and billed at Flash pricing, until a V4.1 Pro model is ready.
That sequencing is unusual. AI labs typically retire a smaller model when a bigger one supersedes it. DeepSeek did the opposite, and the technical report gives a specific reason: memory has become the binding constraint on serving these models cheaply, ahead of raw parameter count.
An encoder-decoder split built for cache compression
Most large language models released since 2023 share a decoder-only design, where every layer recomputes and stores its own key-value cache as it reads through a prompt. DeepSeek's technical report describes V4.1 Flash using a different structure it calls a Causal Encoder-Decoder: a 40-layer transformer split into a 20-layer encoder and a 20-layer decoder. In this design, the decoder's key-value cache is projected once from the encoder's final hidden states rather than rebuilt independently at every decoder layer. That single change is what lets the model activate just 8 billion parameters during prompt processing and 16 billion during generation, despite the 552 billion parameter backbone sitting behind both stages.
Layered on top of that split, the report describes a second mechanism called Compressed Sparse Attention 2, which assigns each attention layer one of three fixed modes, labeled Full, Reindex, or Reuse, so that most layers share cached key-value data and sparse-attention indices rather than recomputing them. A Hierarchical Sparse Indexer further limits how much of the context window later layers need to search. Combined with 4-bit key-value storage, DeepSeek reports the global cache footprint drops to 890 bytes per token, about a quarter of what V4-Flash required and roughly 437 times smaller than DeepSeek's original 2023 model, according to the figure published in the technical report.
The company frames the payoff in blunt terms: a quarter of the high-bandwidth memory and an eighth of the SSD storage per token compared with the prior Flash generation. For any operator running agentic workloads, where the model reads far more tokens than it writes, cache-hit costs make up a large share of the bill. Shrinking the cache shrinks that bill directly.
Where the benchmarks hold up, and where they don't
DeepSeek's own reported numbers, evaluated internally, show V4.1 Flash ahead of V4 Pro on Terminal-Bench 2.1 (90.6 versus 87.9), DeepSWE v1.1 (74.2 versus 62.7), CyberGym (88.1 versus 83.3), and AutomationBench (54.8 versus 43.2). Against outside frontier models, the comparison is mixed rather than one-sided. V4.1 Flash reports a Terminal-Bench 2.1 score of 90.6, ahead of the figures DeepSeek lists for GPT-5.6 Sol (88.8) and Kimi K3 (88.3).
On Humanity's Last Exam, though, V4.1 Flash trails badly, scoring 36.8 against the 56.3 DeepSeek attributes to Claude Opus 5. On the base model's world-knowledge benchmarks, SimpleQA-Verified included, V4.1 Flash also falls well behind V4 Pro, 42.3 versus 55.2, a reminder that the parameter cuts driving the cost savings show up as real capability tradeoffs on tasks that depend on stored knowledge rather than reasoning steps.
BenchmarkV4 ProV4.1 FlashTerminal-Bench 2.1 (Pass@1)87.990.6DeepSWE v1.1 (Resolved)62.774.2CyberGym (Pass@1)83.388.1AutomationBench (Pass@1)43.254.8SimpleQA-Verified, base model (EM)55.242.3Humanity's Last Exam (Pass@1)42.736.8
The technical report also discloses that results shift depending on which agent scaffold runs the evaluation. On DeepSWE v1.1, the model scores 74.2 under the mini-SWE harness but only 67.6 under DeepSeek's own "PTC" harness variant, a seven-point spread from swapping the surrounding tooling rather than the model itself. That detail matters for anyone comparing benchmark tables across labs, since scaffold choice can move a headline number by more than the gap between competing models.
The business logic behind retiring V4 Pro
DeepSeek's stated reason for phasing out V4 Pro is that internal and external testing put V4.1 Flash ahead of it on performance, cost, speed, and total completion time. Whether or not every one of those claims holds up under independent testing, the pricing move is concrete and verifiable through the API itself: V4 Pro requests get rerouted to Flash-tier pricing on September 14, and DeepSeek has already cut Flash-series prices, with off-peak rates set at half of peak rates as of September 10.
The timing lands amid a broader repricing fight among Chinese AI labs, where chip export restrictions have tightened the compute available to train and serve models, pushing developers toward efficiency gains that substitute for raw hardware access. It also lands as DeepSeek is reportedly preparing for a listing on Shanghai's STAR Market, having engaged CITIC Securities among four underwriters according to Reuters reporting relayed by the South China Morning Post, alongside a pre-IPO financing round that could value the company near 500 billion yuan. A public listing raises the stakes on demonstrating that DeepSeek's efficiency claims translate into a durable cost advantage rather than a one-time architectural trick.
Reading the cache math against the on-prem argument
The specific number worth tracking is 890 bytes per token. That figure sets the memory budget for anyone considering running this model outside a hosted API, on their own GPUs rather than through DeepSeek's cloud endpoint. Key-value cache size is one of the main variables that determines how much local hardware a self-hosted deployment needs, alongside the model's raw parameter count. A model that needs a quarter of the cache memory to serve the same context length changes that math meaningfully for smaller deployments weighing a local instance against a subscription API. Whether that shift moves the inflection point on the on-premises calculation depends on how these compression techniques hold up once independent labs run their own memory benchmarks rather than relying on the vendor's reported figures, and that verification hasn't happened yet.
What to watch next
The open weights are already on Hugging Face under an MIT license, and DeepSeek says it plans to work with the open-source inference community on deployment support. That means the compression claims are checkable in a way marketing claims usually aren't. Anyone running the model on their own hardware within the next few weeks will be able to measure the actual memory footprint directly rather than taking DeepSeek's figure on faith, and the scaffold-dependent variance disclosed in the technical report suggests the agentic benchmark gap will need retesting outside DeepSeek's own harness before the comparison to V4 Pro, GPT-5.6 Sol, and Claude Opus 5 can be called settled.
Author Bio
David Borish is a writer covering frontier AI development and enterprise deployment economics. He publishes The AI Spectator and is the author of The Tony Hawk Paradox: When Video Games Predict Reality. More of his work is available at davidborish.com
