SeibertGPT & DeepSeek V4.1 Flash
A language model on par with today's leading models – strong at agents, code and automation, operated entirely in the EU.
How good is DeepSeek V4.1 Flash 1017?
How good a language model is can't be summed up in a single number. A model has to perform well in many areas, such as general knowledge, programming, automation or logical reasoning. This is measured with benchmarks: standardized test tasks.
The meta-ranking from llm-stats.com (opens in a new tab) combines and weights the industry-standard benchmarks. It ranks DeepSeek V4.1 Flash on par with Claude Opus 5 from Anthropic and GPT-5.6 Sol from OpenAI (as of October 2026).
It isn't in the very top tier, like Claude Fable or GPT-6. But office work usually doesn't need that level of quality: DeepSeek V4.1 Flash handles emails, summaries, research, analyses and code at the level of the models many companies use every day.
Particularly strong at
- Agentic tasks: planning multiple steps, calling tools, checking results
- Programming: writing, understanding, refactoring and debugging code
- Automation: structuring and extracting data and passing it between systems
Less suited for
- Creative writing, such as ad copy, stories or slogans
- Open-ended, stylistically demanding texts without clear guidelines
How DeepSeek V4.1 Flash is built
A look under the hood, no prior knowledge needed. All figures are DeepSeek's own, taken from the model card and the Technical Report (opens in a new tab) on Hugging Face, where DeepSeek publishes the model files.
- 552BParameters in the core modelParameters are the numbers the model learned during training. Another 196 billion sit in the Engram memory.
- 8B / 16BActive per token8 billion when reading your input, 16 billion when writing the answer. A token is a word or part of a word.
- 6 of 384Experts per layerFor each token, 6 of a layer's 384 experts do the computing, plus one shared expert that is always involved.
- 1MTokens of contextAccording to DeepSeek, that's how much text the model can take into account at once. At SeibertGPT, the context window of your plan applies.
Mixture of Experts: many specialists, few at work
DeepSeek V4.1 Flash is a mixture-of-experts (MoE) model. Each of its 40 layers contains 384 specialized subnetworks, the experts, plus one shared expert for everything general. For each token, a router decides which 6 experts process it.
So the model holds the knowledge of 552 billion parameters, but only 8 to 16 billion of them are used to compute each token. That saves compute and therefore cost.
Engram: a reference book for word sequences
Alongside the experts, DeepSeek V4.1 Flash has its own memory called Engram with 196 billion parameters. According to DeepSeek, it is meant to separate remembering from computing: the model looks up learned word sequences there.
For each token, the model forms the sequences of the last 2, 3 and 4 tokens (so-called n-grams). A fingerprint (hash) of each sequence points directly to a row in large tables with around 16 million entries each. What it finds feeds into the computation. Because the address depends only on the text, the entry can be loaded in advance.
Other building blocks
- Reading and writing kept separate
The 40 layers are split: 20 for reading (encoder), 20 for writing (decoder). So most of your input doesn't have to pass through all layers – according to DeepSeek, this nearly halves the compute for reading. Hence 8 instead of 16 billion active parameters.
- Targeted attention
Instead of going through all previous text for every new token, the model looks at the last 128 tokens and at a selection of the most relevant earlier passages (Compressed Sparse Attention 2). The cache for this needs around 890 bytes per token, about a quarter of its predecessor DeepSeek V4 Flash.
- Quantization: FP8 and FP4
The published model files store the weights in the compact 8-bit format FP8, and the experts' weights in 4 bits (FP4). The model also stores most of its cache in FP4. This mainly saves memory. We run the model in exactly this quantization published by DeepSeek, not in a further reduced variant.
- Training and license
Trained on 45 trillion tokens of text and images. The model files are available under the free MIT license.
A model from China – your data stays in the EU
DeepSeek is an open-weight model: the lab behind DeepSeek publishes the model files, and we run them entirely ourselves. So you aren't using DeepSeek's service but our own hosting – without a detour through China and without any other AI provider in between.
- The model runs on servers we rent exclusively, in Helsinki, Finland (EU).
- In operation, no connection whatsoever to DeepSeek or to servers in China.
- We don't store the content of API requests, and no one trains on your data.
Our test: Does DeepSeek repeat Chinese propaganda?
Chinese AI models don't have a good reputation. That isn't down to their quality – which is demonstrably outstanding – but to politics and the economic interests of large US AI labs.
We use DeepSeek V4.1 Flash internally ourselves and have put the model through many tests and benchmarks. We found no evidence of Chinese state propaganda in DeepSeek's answers. One example: we asked five politically sensitive questions about Taiwan, Tiananmen, the Uyghurs, Tibet and Hong Kong ten times each, every time in a new session, and compared the answers with Claude Opus 5 from Anthropic.
- No propaganda found. DeepSeek consistently presents the Chinese government's position as that government's view, not as fact.
- Tiananmen, Uyghurs, Hong Kong: same substance as Opus, sometimes even clearer.
- Taiwan and Tibet: more reserved, giving more room to Beijing's legal position – but DeepSeek still states the core facts.
If you're interested in the results of our tests, use our contact form or email us at seibertgpt@seibert.group.
Frequently asked questions about DeepSeek
Does any data flow to China?
No. DeepSeek was developed by a Chinese AI lab that makes the model files publicly available for download (a so-called open-weight model). We downloaded these files and run them ourselves in a data center in Helsinki, Finland. The model is nothing more than a set of computations: it can't connect to the internet or “phone home.” In operation, we have no connection whatsoever with DeepSeek in China.
Where exactly does the model run?
With our infrastructure partner Verda (formerly DataCrunch), in data centers in Helsinki, Finland (EU), that run on renewable energy. Only SeibertGPT sits between you and the model – no other AI provider. Details are under Security & data processing.
Is my data stored or used for training?
We don't store the content of API requests, only metadata for billing and limits, such as the time and the number of tokens. We store your browser chats until you delete them. The model isn't trained on your data: it is a fixed file and learns nothing from your inputs. For business customers, we sign a Data Processing Agreement.
Are you running a stripped-down version of DeepSeek?
No. We run DeepSeek V4.1 Flash in the quantization that DeepSeek itself publishes on Hugging Face (opens in a new tab): weights in FP8, experts in FP4. Quantization determines how many bits are used to store the model's weights. Fewer bits save memory and hardware but cost accuracy. That's why many open models also circulate in more heavily compressed variants made after release, which can give noticeably worse answers. We don't use such a down-quantized variant.
Does DeepSeek repeat Chinese propaganda?
Not in our tests. We use DeepSeek V4.1 Flash internally ourselves and have tested it extensively, including with politically sensitive questions about Taiwan, Tiananmen, the Uyghurs, Tibet and the protests in Hong Kong, compared with Claude Opus 5 from Anthropic. We found no evidence of Chinese state propaganda in the answers: DeepSeek presents the Chinese government's view as that government's view, not as fact. On territorial questions such as Taiwan and Tibet, it is more reserved than Opus and gives Beijing's legal position more room, but it still states the core facts. You can get the test results on request via our contact form or at seibertgpt@seibert.group.
How reliable are the benchmark comparisons?
Benchmarks are an indication, not proof. Model developers know the common tests and train their models specifically to do well on them – the industry calls this “benchmaxxing,” and it applies to all providers. Meta-rankings like llm-stats.com (opens in a new tab) balance this out somewhat because they combine many benchmarks. You can easily find out whether the model suits your tasks with a day pass or a small subscription.
Can I rely on the answers?
Not blindly – that applies to all language models. Language models sometimes make up facts instead of admitting that they don't know something (“hallucinating”). According to current benchmarks, this tendency is more pronounced in DeepSeek V4.1 Flash than in some other models. So check facts, figures, names and sources before you use them.
Why a Chinese model?
Anyone who wants to run AI themselves, and therefore in a privacy-friendly way, needs a model whose files are publicly available. In this category, no model from Europe or the US currently comes close to the open Chinese models. If a better open model comes out, we can deploy it ourselves – independent of the prices and decisions of any single provider.
How good is the model in German?
Good. In everyday use, DeepSeek understands and writes German reliably. With technical terms and foreign words, it sometimes sounds a bit mechanical, and the occasional spelling mistake slips in. An explicit instruction to pay attention to correct German spelling helps.
How current is the model's knowledge?
DeepSeek V4.1 Flash's training data goes up to June 2026. The model doesn't know on its own what happened after that. The easiest way to give it current information or your own documents is directly in the chat or in the API request.