EUdrift · DeepSeek V4.1-Flash hosted in the EU

How a request is handled. Nothing is kept beyond the cache.

The request is processed in the EU by the original model. The response returns to you. The cache is time-limited and then deleted.

Fig. 1 · Request pathSchematic

Your app Calls the API.

EUdrift API Receives the request in the EU.

EU Everything inside the line happens in the EU.

DeepSeek V4.1-Flash Original weights, rev. 1017. FP8 with FP4 experts.

Cache Time-limited cache in the EU.

Deleted Nothing remains.

The model. What it is and how it measures up.

Model details and publisher-reported results from DeepSeek's model card, with source and date. Our own measurements will be published once the service runs.

Overview

Context 1Mtokens Per the model card

Architecture MoE Mixture-of-Experts

Licence MIT Open weights

Revision 1017Versioned Updated with the model

Weights: Original, without additional quantization
Precision: FP8, FP4 expertsPer the model card config
Inference: DeepInferDeepSeek’s original inference software
Hosting: EU

Model ID

deepseek-flash

Sources

Model card huggingface.co

Benchmarks

Show as table Show as chart

Reasoning · Publisher-reported results
Benchmark	Metric	DS-1017	DS-V4.1	K3	GLM-5.3	GPT-5.6 Sol	Opus-5.0
HLE w/o tools † Text-only subset of HLE; not directly comparable with the full test.	Score	42.6	36.8	43.5	42.0†	44.5	56.3
Apex Bench v1.1	Score	72.1	64.6	47.8	56.6	54.3	65.8

Agentic · Publisher-reported results
Benchmark	Metric	DS-1017	DS-V4.1	K3	GLM-5.3	GPT-5.6 Sol	Opus-5.0
Terminal-Bench 4.0	Score	42.3	31.2	12.6	37.9	39.9	51.8
DeepSWE v1.1	Score	77.0	74.2	67.5	66.9	73.0	74.0
ProgramBench	Score	29.1	20.3	17.5	19.0	23.0	37.0
NL2Repo-Bench	Score	68.2	65.4	58.0	58.0	56.8	75.3
Automation-Bench v1.0.6	Score	60.5	54.8	46.7	48.8	45.8	50.3
Agents’ Last Exam	Score	35.4	31.8	27.6	28.5	26.7	28.6
Chartography (w/tools)	Score	83.0	78.9	68.1	n/a	79.9	84.0
BabyVision (w/tools)	Score	92.5	89.6	85.7	n/a	88.9	94.1
ZeroBench-main (w/tools)	Score	51.0	49.0	41.0	n/a	53.0	52.0

Reasoning Publisher-reported results

DS‑1017 Compared models

Publisher-reported results

HLE w/o tools Score

† Text-only subset of HLE; not directly comparable with the full test.

Apex Bench v1.1 Score

Agentic Publisher-reported results

DS‑1017 Earlier DeepSeek (V4.1) Other vendors (K3, GLM‑5.3, GPT‑5.6 Sol, Opus‑5.0)

Publisher-reported results

Publisher comparison for revision 1017. Terminal-Bench uses Claude Code; DeepSWE uses Mini-SWE-V2. † marks the text-only HLE subset. Model card, retrieved 2026-10-22.

Publisher's figures

KV-cache. Illustrative

KV cache and relative comparisons are published once the model card data has been verified for revision 1017.

Chart awaiting verified data

Independent measurements

Independent measurements.

We link to the source instead of embedding data. Real values will be published with source and date once reuse terms are agreed.

Chart awaiting verified data

Specification

Architecture: Mixture-of-Experts
Context: 1M tokens
Layers: 40
Experts: 384 routed, 6 active1 shared
Tensor types: FP8 (F8_E4M3), FP4 expertsPer config.json
Licence: MIT

Details come from DeepSeek's model card and config for revision 1017. Publisher figures are not Eudrift measurements.

DeepSeek V4.1-Flash-1017 model card

Data handling

Region: EU
Training on your data: No
Content logging: No
Cache: Time-limited, in the EU.
DPA: Included
Subprocessors: Published before access opens.

In short

Request content is not stored beyond the time-limited cache. Your data is not used for training.

How a request is handledFig. 1

Our ambition. Four things we build for.

Goals for the service, not measurements. Measurable levels are published once verified.

01

Original weights

No additional quantization.

02

High availability

A goal for operations from day one.

03

Effective caching

Lower cost when prefixes are reused.

04

Hosted in the EU

With a DPA and clear terms.

Indicative prices. Per million tokens.

Indicative API pricing. Register your interest — no commitment.

€0.006/M

Cached input

€0.25/M

Uncached input

€1.20/M

Output

Indicative prices excl. VAT. Registration is free and non-binding.

How caching saves. Same start, lower price.

When a request starts like an earlier one, for example with the same system prompt or document, that part is reused. It is billed as cached input and the response arrives sooner.

The cache is in the EU, time-limited and deleted after expiry. How much is reused depends on your traffic.

Request 1

System prompt + documentQuestion

Request 2

Same start, cachedNew question

Questions. Short answers.

Answers about the service.

Where is data processed?

Inference and caching take place in the EU. Request content is not stored beyond the time-limited cache.

What precision does the model run at?

The model card for revision 1017 states FP8 with FP4 experts. We will use the original published weights without additional quantization, with DeepSeek’s own DeepInfer inference software.

How does the cache work?

Repeated request beginnings are reused and billed as cached input. The cache is time-limited, located in the EU and deleted after expiry. The cache hit rate depends on your traffic.

DPA and subprocessors?

The service will include a data processing agreement (DPA). The subprocessor list will be available before access opens.

Is the API compatible?

Yes. The service will provide the standard OpenAI-compatible Chat Completions endpoint, /v1/chat/completions. Use the OpenAI SDK with our base URL and API key, and deepseek-flash as the model ID.

How high is availability?

High availability is a goal for operations. Measurable levels are published once set and verified.

What happens after I register interest?

We get in touch when access opens, or call you if you asked us to. No account is created and nothing is bought.

DeepSeek V4.1-Flash. Hosted in the EU.