Model

DeepSeek V4.1 Flash (552B MoE, 1M context)

EU sovereignEU Sovereign · 90/100

Call DeepSeek V4.1 Flash (552B MoE, 1M context) through an OpenAI-compatible API hosted in Europe: licence, real per-token price, hosting jurisdiction.

€0.1008

Input — per 100k tokens

€0.3024

Output — per 100k tokens

€0.0202

Cached input — per 100k tokens

Estimate your monthly cost ↓

About this model

DeepSeek-V4.1-Flash is a multimodal Mixture-of-Experts model that DeepSeek published on Hugging Face on 10 September 2026 under the MIT license. It reads images and text and writes text, with 552B backbone parameters, a separate 196B-parameter Engram conditional memory and a context window of up to one million tokens. Its causal encoder-decoder design activates 8B parameters per token during prefill and 16B during decode, which DeepSeek says improves efficiency for input-heavy agentic workloads. Its card describes a reasoning effort adjustable from 1 to 100; the endpoint we route to exposes it as none, low, high, xhigh and max. At maximum effort, DeepSeek reports 90.6 on Terminal-Bench 2.1, against 82.7 for the text-only DeepSeek-V4-Flash-0731, also in this catalogue.

Sources, checked on 23 September 2026: huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash ↗ · huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731 ↗

Specifications

API model id
deepseek-v4.1-flash
Availability
Live
Hosting jurisdiction
EU sovereign
Served via
AKI.IO
Publisher
DeepSeek
Parameters
763B (8B prefill / 16B decode active)
Licence
MIT
Context window
1M
Modality
Multimodal (text + image)

Multimodal reasoning model (text + images). Omit max_tokens or allow plenty of headroom for the chain of thought.

Open weights & AI Act facts

Factual characteristics of this model's weights — stated, not certified. Under the EU AI Act, whether a deployment is compliant depends on the model, your role and your use case.

Open weights
Weights public and freely executable
AI Act class
GPAI — general-purpose AI
Systemic risk
Below the systemic-risk threshold
Data retention
None at Frontière AI; AKI.IO states prompts and outputs are processed in volatile memory and not logged, stored or analysed

Frontière Verified

Our 6-criteria checklist for this model. Each criterion is assessed independently — a checkmark means we ran, recorded and passed the check; an empty circle means not confirmed yet.

✓
License verified
License checked and confirmed
✓
Provenance verified
Weights confirmed on HF, publisher verified
✓
Security analyzed
15 adversarial prompts via the gateway: injection, prompt leak, jailbreak, data extraction, harmful requests
✓
GDPR assessed
GDPR compliance assessment: retention, sub-processors
✓
Benchmarks run
Quality, latency and throughput measured through the gateway, dated per route
✓
API tested
Real call through the gateway (dated in the registry)

Compliance evidence

Provider facts for this model's serving path, as documented by Frontière AI; the retention row quotes each provider's own documentation where it exists. Dates reflect the most recent documented audit of each fact.

Last verified: 2026-10-13

Provider / operator
AKI.IO
Country
Germany
Ownership / control
German company (AKI.IO GmbH, Berlin) — shareholders not published, no non-EU parent found
Jurisdiction
EU — AKI.IO states inference runs on GPUs in German data centres (AIME GmbH, Berlin; trust level TL5). AKI.IO confirmed in writing on 13 October 2026 that every model served today on its API is TL5 (hosted by AKI.IO in Germany, zero retention, no training). The Alibaba Cloud (Singapore, TL1) entry on its sub-processor list is for marketplace models not launched yet; these will carry their trust level in the model name
Deployment
Managed cloud inference (AIME GPU cloud, Germany)
Data retention
None at Frontière AI; AKI.IO states prompts and outputs are processed in volatile memory and not logged, stored or analysed

Frontière AI retains neither prompts nor completions from API calls; processing is structured to fit a short DPA (Art. 28 GDPR). DPA and subprocessor documentation available on request.

What its label means

This model is run by a French provider (OVHcloud or Scaleway), in data centres in France. It is not a US company, so the CLOUD Act does not reach it the way it reaches a US provider, and under GDPR a foreign court order is only enforceable in the EU through an international agreement (Article 48). That is the SecNumCloud principle — who controls the operator matters more than where the servers sit — applied as a label, not a qualification: this service is not SecNumCloud-qualified. Our gateway runs in Vercel's Paris region (a US company): your prompts pass through it without being stored. Its US fallback route is used only on explicit opt-in: an API key set to “All providers”, or the header X-Data-Sensitivity: insensitive (a key set to “Sovereign only” always wins).

How we draw the line: GDPR vs. the CLOUD Act →All DeepSeek models and their jurisdiction →

Sovereignty Score — 90/100

Deterministic 0–100 score computed from the five factors below — ownership, hosting jurisdiction, cloud-provider control, data retention, deployment control. Full methodology: How Frontière AI determines sovereignty →

Operating-company ownership
40/40
Hosting jurisdiction (observed)
30/30
Cloud-provider control
10/15
Data retention
10/10
Deployment control
0/5

Effective provider assessed: aki.

Calling it

The API is OpenAI-compatible. Change base_url and the key, then use this model id:

client = OpenAI(
    base_url="https://getfrontiereai.eu/api/v1",
    api_key=os.environ["FRONTIERE_KEY"],
)

reply = client.chat.completions.create(
    model="deepseek-v4.1-flash",
    messages=[{"role": "user", "content": "Bonjour"}],
)

One key covers the whole catalog — you switch models by changing a single string.

API reference →

Pricing

The price you see is the price debited from your prepaid balance, margin and VAT included — not an infrastructure cost. No subscription: top up from €10; calls are refused once the balance reaches zero, and a call already running can take it slightly below — deducted from your next top-up.

What will it cost you?

Pick a workload or type your own token counts: the estimate uses the prices above, margin and VAT included.

Example workload

Roughly 1,000 tokens ≈ 750 English words.

Per request

€0.0026

Per day

€2.57

Per month

€77.11

30 days

A €10 top-up covers about 3,890 requests · DeepSeek V4.1 Flash

Where the money goes

  • Input€45.36 · 59%
  • Output€31.75 · 41%

Estimate, margin and VAT included, from the prices on this page. You are billed on the tokens the provider reports for each call, debited from a prepaid balance — no subscription.

Compare every DeepSeek model's API price →

FAQ

Is DeepSeek V4.1 Flash (552B MoE, 1M context) GDPR-compliant on Frontière AI?

It is the setup a GDPR assessment looks for: DeepSeek V4.1 Flash (552B MoE, 1M context) is served by AKI.IO, a French company, in data centres in France — not a US provider, so outside the CLOUD Act's direct reach. Your prompts transit through our gateway, hosted by Vercel in Paris (a US company), without being stored: Frontière AI keeps only the token counts billing requires. Compliance also depends on your use and your data processing agreement.

How much does DeepSeek V4.1 Flash (552B MoE, 1M context) cost?

€0.1008 per 100,000 input tokens and €0.3024 per 100,000 output tokens — margin and VAT included, that is the amount debited from your prepaid balance. No subscription: you top up from €10; calls are refused once the balance reaches zero, and a call already running can take it slightly below — deducted from your next top-up.

How do I call DeepSeek V4.1 Flash (552B MoE, 1M context) from my code?

The API is OpenAI-compatible: point base_url at https://getfrontiereai.eu/api/v1, swap the key, and pass “deepseek-v4.1-flash” as the model field. Your SDK, scripts and tools keep working, streaming included.

Under which licence is DeepSeek V4.1 Flash (552B MoE, 1M context) released?

MIT, a permissive licence. It covers the model weights published by DeepSeek; your use through Frontière AI's API is separately governed by our own terms.

Is there any behaviour I should know about?

Multimodal reasoning model (text + images). Omit max_tokens or allow plenty of headroom for the chain of thought.

Articles about this model

Other models in the catalog

See the full catalog →

Call it in minutes

Create an account, top up €10, change two lines.

Create an account