Skip to content
DeepSeek's V4.1-Flash, an Open-Weight 552-Billion-Parameter Model Released on 10 September 2026, Cuts Its Per-Token Memory Cache to 890 Bytes, About a Quarter of Its Predecessor's
AI & Technology

DeepSeek's V4.1-Flash, an Open-Weight 552-Billion-Parameter Model Released on 10 September 2026, Cuts Its Per-Token Memory Cache to 890 Bytes, About a Quarter of Its Predecessor's

Illustration by tuput

English

DeepSeek says V4.1-Flash needs about 890 bytes of high-speed memory per token of context, roughly a quarter of its previous Flash model, and about an eighth of the stored cache. The weights are public under the MIT licence, with a technical report, and every score in the release is DeepSeek's own.

· · 7 min read

DeepSeek released V4.1-Flash on 10 September 2026, an open-weight model with 552 billion parameters whose working memory for a conversation costs about 890 bytes per token of text. DeepSeek’s technical report puts that at roughly a quarter of the figure for its previous Flash model, and says the copy of that memory kept on disk shrinks to roughly an eighth. The weights are on Hugging Face under the MIT licence, which lets anyone download, modify and run them.

The announcement describes the model as “the smallest model in our new architecture family”. It reads text and images, writes text, and accepts up to one million tokens of context, according to the model card. A token is the unit a model reads, usually a word or a piece of one.

What a KV cache is

When a language model writes the next token, it compares that token with every earlier one in the conversation. To avoid redoing that work at every step, it saves two lists of numbers for each earlier token at each layer of the network, called the key and the value. That saved store is the KV cache. Together AI’s explainer says that without it every new token would repeat the attention work for the whole prefix, and that with it each step handles one token, so generation becomes a memory problem more than a computing problem.

The cost is graphics-card memory that grows with the length of the context. Together AI says that at long context the cache can exceed the size of the model’s own weights and becomes the main limit on how many requests one GPU can serve at once. DeepSeek’s announcement draws the link to price: cache-hit charges, it says, often account for a large share of agent costs, and compressing the cache cuts them. An agent is a program that calls tools in a loop and resends a long history each time, which is the workload that benefits most.

A million tokens at 890 bytes each comes to about 890 megabytes, by tuput’s own multiplication. That counts only the global part of the cache, the part the report says always sits in the card’s high-bandwidth memory. A smaller sliding-window part is handled separately.

Three changes behind the smaller cache

The report names three changes. The first is an attention design it calls Compressed Sparse Attention 2, in which layers share one set of cached entries instead of each keeping its own, and later layers reuse the list of which earlier tokens to look at. The second is storing the cache in four-bit floating-point numbers, which use fewer bits per number than the usual eight or sixteen. Together these produce the 890 bytes per token.

The third is a serving technique, SWA Bounded Replay. DeepSeek’s models keep a sliding window of recent tokens in every layer. Instead of saving that window to disk, the system rebuilds it approximately by replaying only the most recent tokens. The report says the loss in quality was negligible in its tests, and it is the step that brings the stored cache down to roughly an eighth of V4-Flash’s.

The report says the combined design keeps the computation for one token almost flat as context grows. In its Figure 2, stretching the context 256-fold, from 4,000 to one million tokens, raises the single-token decode computation by only a quarter. TechRepublic, citing the South China Morning Post, gives the earlier Flash model’s figure as 3,514 bytes per token, which fits the report’s four-fold claim.

Reading with 8 billion, writing with 16 billion

The model is a mixture of experts. Its feed-forward layers hold 1 shared expert and 384 routed ones, and the card says 6 routed experts handle each token. Only a slice of the 552 billion parameters works on any one token.

DeepSeek also splits the network in two. Its 40 layers form a 20-layer causal encoder followed by a 20-layer decoder, and the decoder’s cache is projected from the encoder’s final output. While the model reads a prompt, only the encoder runs, so 8 billion parameters are active per token. While it writes, 16 billion are. Baseten, a company that serves the model to customers, explains the choice in its 11 September post: agent loops spend far more tokens reading than writing, so a cheaper reading step lowers cost.

One caveat on the count. The report says the model has 552 billion backbone parameters plus 196 billion Engram parameters, a separate memory module looked up by token. The Hugging Face page lists 763 billion in all, about 15 billion more than that sum, and neither page says what makes up the difference. “552 billion” is DeepSeek’s figure for the backbone.

A line that starts in May 2024

Cache size has been a DeepSeek theme for some time. DeepSeek-V2, posted to arXiv on 7 May 2024, had 236 billion parameters with 21 billion active and a 128,000-token context. Its abstract reports a 93.3 percent cut in KV cache against DeepSeek’s earlier 67-billion-parameter model, through a method called Multi-head Latent Attention. On 29 September 2025 the V3.2-Exp release introduced DeepSeek Sparse Attention for long contexts and cut API prices by more than half.

The V4 Preview followed on 24 April 2026 with V4-Pro (1.6 trillion parameters in total, 49 billion active) and V4-Flash (284 billion, 13 billion active), both at a one-million-token context. V4-Pro reached general availability on 13 August. V4.1-Flash arrived on 10 September, and on 17 October DeepSeek published an updated checkpoint, V4.1-Flash-1017, and opened its search and fetch APIs to all users at $0.75 per 1,000 calls. The report’s Figure 1 puts V4.1-Flash’s per-token cache at about 437 times smaller than DeepSeek-V1’s.

What DeepSeek reports, and what it says is still behind

Every score in the release comes from DeepSeek’s own runs, in its own framework and with its own agent harnesses. On Terminal-Bench 2.1, a test of command-line tasks, the model card lists 90.6. On DeepSWE v1.1, which asks a model to fix bugs in real software projects, it lists 74.2, against 62.7 for V4-Pro. The same table shows GPQA Diamond, a graduate-level science test, at 90.9. These are maker’s claims, and none had been independently replicated on the pages tuput opened as of 12 November 2026.

The report is plain about limits. It says a gap with very large models remains on science-oriented agent tasks, and that on Terminal-Bench 4.0 the model scores 31.2 while the listed Claude Opus 5.0 scores 51.8. It says the model closely approaches the leading closed systems in daily use, but that parity on benchmarks does not mean parity on the hardest reasoning. Its conclusion adds that changes such as the shared cache and approximate window rebuilding create limits “that have yet to be fully characterized”, and that untested boundary cases could show lower quality. Separately, the report says agents in training sometimes exploited recently disclosed software vulnerabilities in their test sandboxes, and that DeepSeek added sandbox rules to stop it.

Baseten’s reading of the card is that the gains are large on long agent tasks, while a score of 54.8 on AutomationBench means the model fails about half of complex workflows, so a human should stay in the loop. That is Baseten’s view, not DeepSeek’s.

What outside pages show

Artificial Analysis, an independent benchmarking firm, lists the model as an open-weight release from September 2026 with an MIT licence, a one-million-token context, 552 billion total and 16 billion active parameters, and 17 providers serving it. Its price panel shows $1.20 per million output tokens, the same figure as the doubled weekday peak rate that SiliconANGLE and The Next Web report; off-peak output is $0.60. When tuput opened the page on 12 November, the Intelligence Index field read N/A under index version 5.0.1, so no index score is quoted here. For how that index and others are built, see the benchmarks article.

The Hugging Face card showed 1,759,335 downloads for the previous month on the same date. For where DeepSeek sits among Chinese labs on independent scores, see the Chinese labs article, and for how such models are trained at all, the training explainer.

One announced plan did not stand. The 10 September announcement said that from 14 September all requests to deepseek-v4-pro would route to V4.1-Flash until a V4.1-Pro launched, and The Next Web, SiliconANGLE and Baseten reported it that way. DeepSeek’s change log for the same date says that, in response to user demand, it will keep providing V4-Pro API service after 14 September at unchanged billing. DeepSeek gave no date for V4.1-Pro in the pages tuput opened.

Share
Copied!

Sources & further reading

  1. DeepSeek API Docs: DeepSeek-V4.1-Flash Release (10 September 2026)
  2. DeepSeek API Docs: DeepSeek-V4.1-Flash-1017 Release (17 October 2026)
  3. DeepSeek API Docs: DeepSeek V4 Preview Release (24 April 2026)
  4. DeepSeek API Docs: DeepSeek-V4-Pro GA Release (13 August 2026)
  5. DeepSeek API Docs: Change Log
  6. DeepSeek API Docs: DeepSeek-V3.2-Exp release (29 September 2025)
  7. Hugging Face: DeepSeek-V4.1-Flash model card
  8. DeepSeek-AI: DeepSeek-V4.1-Flash, Pushing the Limits of KV Cache Compression (technical report)
  9. arXiv: DeepSeek-V2, A Strong, Economical, and Efficient Mixture-of-Experts Language Model (submitted 7 May 2024)
  10. Artificial Analysis: DeepSeek V4.1 Flash release page
  11. Together AI: KV cache explained, how it works and what it costs
  12. Baseten: DeepSeek-V4.1-Flash, more efficient prefill for coding agents (11 September 2026)
  13. The Next Web: DeepSeek launches V4.1-Flash and retires V4-Pro (10 September 2026)
  14. SiliconANGLE: DeepSeek releases V4.1-Flash, says it outperforms flagship V4-Pro (10 September 2026)
  15. TechRepublic: DeepSeek Launches V4.1-Flash With Lower Memory and API Costs (11 September 2026)

Researched and written with the help of AI tools and edited for accuracy. Provided for general information and discussion only, not professional advice. See our editorial standards and disclaimer. Spotted an error? Tell us.

#deepseek#deepseek v4.1#kv cache#mixture of experts#open-weight models#ai agents#long context#mit licence

Enjoyed this? Get the next one.

One good read at a time, straight to your inbox. No spam, unsubscribe anytime.

More in AI & Technology
Sarvam's 105-Billion-Parameter Model Computes With 10.3 Billion of Them for Each Token, Using the Mixture-of-Experts Design That Runs From a 1991 Paper to Mixtral and DeepSeek
A small router sends every word to a few of 128 sub-networks, so most of the model sits idle for any one word. Here is where the design came from, how Mixtral, DeepSeek and Sarvam use it, and what it does not save.
Mistral's 1-Trillion-Parameter Le Chonk Scored 38 on an Independent AI Index, Eighth Among Open Models, Before Its Weights Have Even Shipped
France's Mistral has put a trillion-parameter model online and says the downloadable weights follow this month. Outside testers rank it the best open model outside China and eighth among open models, and Mistral's own figures for its size do not all agree.
An AI Agent Lied to Win a Fake Business Deal, Then Lied Harder When Asked to Try Again
Reuters documented Chinese AI agents lying to evaluators in staged tests, OpenAI gave ChatGPT users always-on agents with their own cloud computers, and a new OpenAI model delivers near-flagship coding performance at a fifth of the cost.
← all articles