DeepSeek V4.1 Flash explained: how an 8B-active encoder-decoder runs a 552B open model, and what it means before you route work to it

DeepSeek's release notes give you charts and a model ID. This page explains why the model reads with 8 billion parameters and writes with 16 billion, what an 890-byte KV cache means at a million tokens, which benchmark numbers are DeepSeek's own, and why the release landed two days after US agencies named DeepSeek for distillation.

What 'Mixture of Experts' and 'active parameters' mean

A dense model uses every one of its weights on every token. A Mixture of Experts (MoE) model splits most of its layers into many small expert networks and a router picks a few for each token. V4.1-Flash has one shared expert plus 384 routed experts in each MoE layer, and the router turns on 6 routed experts per token. So the 552B number is the size of the whole library on disk. The active number is how much of it is read for any single token, and that is what sets speed and compute cost.

You still need the whole 552B in memory to serve it, because any expert might be chosen next. That is why 'open weights' and 'runs on my laptop' are different claims. At 8 bits per weight the weights alone are about 550 GB; at 4 bits about 275 GB. That is arithmetic, not a DeepSeek figure, and the model card does not state a recommended hardware setup.

The Causal Encoder-Decoder, and why input and output use different amounts of the model

Most chat models are decoder-only: one stack of layers does both jobs, reading the prompt and generating the answer. V4.1-Flash is a Causal Encoder-Decoder (CED): a 40-layer transformer split into a 20-layer causal encoder that processes the input, followed by a 20-layer decoder that produces the output. 'Causal' here means the encoder still only looks backward, so it can stream and cache like a normal decoder.

The payoff is that the two halves can be sized separately. Reading the prompt (prefill) activates 8B parameters per token. Writing the answer (decode) activates 16B. Agent workloads are input-heavy: a coding agent might read 200,000 tokens of repo and tool output to write 2,000 tokens of patch. Making the reading half cheaper is where the cost saving lands for exactly that kind of job. It is also why DeepSeek can claim it is ahead of V4-Pro on 'total runtime,' not just per-token price.

890 bytes per token: what the KV cache number means at a million tokens

When a model reads a long input, it stores a key and a value vector for every token so it does not recompute them on every step. That store is the KV cache, and for long contexts it, not the weights, is what fills up GPU memory. The model card gives V4.1-Flash's global KV cache as 890 bytes per token, stored in FP4 (a 4-bit floating point format, E2M1).

890 bytes times one million tokens is about 0.89 GB for a full-length context. The release notes put it another way: a quarter of the HBM (high-bandwidth GPU memory) and an eighth of the SSD storage of the previous generation. This is the single number that makes a 1M-token window affordable to serve rather than a marketing line. The card says the model was pretrained on 45 trillion tokens, with sparse attention at 64K and the 1M extension trained on 34 trillion of them.

The benchmarks, and whose numbers they are

From the model card: DeepSWE v1.1 74.2 percent (V4-Pro 62.7, GPT-5.6 73.0, Kimi K3 67.5); Terminal-Bench 2.1 90.6 percent (V4-Pro 87.9, GPT-5.6 88.8, Kimi K3 88.3); HumanEval 79.4 percent pass@1. The release notes say 'tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed and total runtime,' and do not name the parties.

Two things to hold in mind. First, every number above was produced or chosen by DeepSeek; independent runs were not available at publish time. Second, the model has a 'continuously controllable reasoning effort' dial from 1 to 100, and the card does not say which setting produced each score, so your cost at that quality is not yet known. Speed claims also lean on DSpark, DeepSeek's speculative decoding, which is a serving trick, not a model property, so self-hosters may not see the same throughput.

What changes on the API on September 14

The model ID is deepseek-flash and it went live September 10 with native image input. New pricing took effect the same day; the exact per-token figures are published as an image in the release notes, with off-peak rates at 50 percent of peak. V4-Pro is being phased out: from September 14, V4-Pro requests are routed and billed at V4.1-Flash pricing. If you have V4-Pro hardcoded anywhere, that is a behavior change with a date on it.

The part the release notes leave out

On September 8, the NSA, CISA and FBI published advisory AA26-251A naming DeepSeek, along with five other Chinese companies, for 'aggressive, malicious, and targeted distillation activities at an industrial scale' against US frontier models. On September 9, China's commerce ministry called distillation a neutral technique and promised countermeasures if the US acts. On September 10, DeepSeek shipped V4.1-Flash under MIT. None of the three documents mentions the others.

The advisory does not say DeepSeek's outputs are unsafe or that using the model is prohibited. It does raise two questions this release cannot answer: how much of the model's skill came from other companies' models, and what happens to prompts sent to DeepSeek's hosted API. Open weights answer the second question for anyone who self-hosts, which at 552B is a data-center decision, not a desktop one.

Who is affected

What to do

• If you use deepseek-flash or V4-Pro through the API, run your evaluation set against V4.1-Flash this week; billing changes September 14.
• Read the pricing image on the release page and write the numbers down. They are not in the page text, so they will not show up in a search later.
• For long-context jobs, test at your real input length. The 1M window is trained, but the card does not publish quality at the far end.
• Before you route client data to the hosted API, decide whether a model named in a US federal advisory is acceptable under your own policy. Write the decision down either way.
• If you self-host, start from the Hugging Face card, not third-party quantized uploads, until the community has vetted them.

What is still unknown

• Independent benchmark results. Every score above is from DeepSeek's model card, and the 'multiple parties' in the release notes are not named.
• The exact API prices. They are published as an image, not text, on the release page.
• Which reasoning-effort setting produced each benchmark score, and therefore the real cost at that quality.
• How much of the model's capability came from distilling other companies' models. The US advisory alleges it; DeepSeek has not responded in the release.
• The technical report's detailed training data description. The card gives token counts, not sources.
• Whether V4.1-Pro, which the release does not mention, is coming, and when.

Sources

• DeepSeek API Docs, 'DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient' — the original report
• DeepSeek-V4.1-Flash model card on Hugging Face (MIT license, architecture, benchmarks)
• DeepSeek launch thread on X (September 10)