DeepSeek is a Chinese AI company that publishes powerful language models mostly as open weights and additionally offers them through its own, very inexpensive API. The current reference is the V4 generation (released 24 April 2026) with 1 million tokens of context: DeepSeek V4.1-Flash (deepseek-flash, since 10 September 2026, currently in the 1017 build of 17 October 2026) as a small, fast and natively multimodal model, and DeepSeek V4-Pro (deepseek-v4-pro, GA build 0813) for demanding agentic and coding workflows. Predecessors such as V3.2, V3.1 and R1 remain relevant mainly via hyperscalers and self-hosting.
For enterprises in the DACH region, the hurdle is less model performance than data residency: the direct API processes data in China, and the hyperscalers offer genuine EU data residency for DeepSeek only in a few configurations. As an AI consultancy in Rosenheim, Germany, we advise on all deployment options, from self-hosting to integration via CompanyGPT. This overview reflects the state as of October 2026.
Current models at a glance
| Model | API ID | Release | Context | Price (input cache miss / output, peak) | Focus |
|---|---|---|---|---|---|
| DeepSeek V4.1-Flash-1017 | deepseek-flash | 17 Oct 2026 (V4.1-Flash: 10 Sep 2026) | 1M | USD 0.30 / 1.20 | Low-cost all-rounder, text and image, agentic coding |
| DeepSeek V4-Pro (0813, GA) | deepseek-v4-pro | 13 Aug 2026 | 1M | USD 1.32 / 3.96 | Agents, coding, complex reasoning |
| DeepSeek V4-Flash-0731 | deepseek-v4-flash (routes to V4.1-Flash) | 31 Jul 2026 | 1M | superseded | Superseded by V4.1-Flash |
| DeepSeek V4-Flash-Vision-Exp | deepseek-v4-flash-vision-exp (routes to V4.1-Flash) | 21 Aug 2026 | 1M | superseded | Experimental, superseded by V4.1-Flash |
| DeepSeek V3.2 / V3.2-Speciale | – | December 2025 | – | – | Hyperscalers (Bedrock, Foundry) |
DeepSeek V4.1-Flash
Per DeepSeek’s changelog, on 10 September 2026 the API model name deepseek-flash was updated to DeepSeek-V4.1-Flash – described as the “smallest model in our new architecture family, with native multimodal visual understanding”. V4.1-Flash replaces V4-Flash (including build 0731) and V4-Flash-Vision-Exp. Image input works via Base64, URL or the Files API; an image is converted to at most 384 tokens depending on its dimensions.
Since 17 October 2026 deepseek-flash has been served by the DeepSeek-V4.1-Flash-1017 build. DeepSeek calls it a “minor update” with substantial improvements in reasoning, coding, agentic tasks and visual benchmarks; per DeepSeek, task runtime and output length also fall, so tasks finish faster and at lower cost. The API call and prices remain unchanged. Per DeepSeek, it was optimised in particular for DeepSeek Harness (recommended: version 0.2.1-rc.1 or newer), Tencent CodeBuddy and WorkBuddy. Benchmarks per DeepSeek, 1017 build versus V4.1-Flash:
| Benchmark | V4.1-Flash-1017 | V4.1-Flash |
|---|---|---|
| Terminal-Bench 4.0 (Claude Code framework) | 42.3 | 31.2 |
| DeepSWE v1.1 (Mini-SWE-V2) | 77.0 | 74.2 |
| NL2Repo-Bench | 68.2 | 65.4 |
| ProgramBench | 29.1 | 20.3 |
| HLE without tools (text subset) | 42.6 (44.0) | 36.8 (39.1) |
| Agents’ Last Exam | 35.4 | 31.8 |
The weights of both builds are available under the MIT license on Hugging Face. Per the model card, V4.1-Flash is a multimodal mixture-of-experts model with 552B backbone parameters, of which 8B are active per token in prefill and 16B in decode, with up to 1M tokens of context; reasoning effort can be set continuously from 1 to 100. The model card also contains a comparison table with competitor models; we do not take any ranking statements from it, since all figures come from the vendor.
DeepSeek V4-Pro
V4-Pro (deepseek-v4-pro, listed in the documentation as DeepSeek-V4-Pro-0813) was released as a preview on 24 April 2026 and reached general availability on 13 August 2026 in the app, web (“Expert Mode”) and API. The model has 1.6 trillion parameters (about 49B active, MoE) and 1M tokens of context at only 27% of the FLOPs and 10% of the KV cache compared with V3.2. Reasoning effort can be set in three levels: low for simple tasks, high for everyday agent work, max for complex tasks; this also applies to Flash. The Responses API format is supported natively (adapted for Codex). Benchmarks per DeepSeek: Terminal-Bench 2.1 87.9, NL2Repo 61.5, DeepSWE 62.7, HLE 42.7 (without tools) / 60.0 (with tools); in April DeepSeek cited 80.6% on SWE-Bench. DeepSeek has confirmed it will continue the V4-Pro API beyond 14 September 2026 on unchanged terms. The weights of DeepSeek-V4-Pro-0813 are available under the MIT license on Hugging Face.
DeepSeek V4-Flash (April base, 0731, Vision-Exp)
The April release V4-Flash (284B parameters, about 13B active, MoE, 1M context, thinking and non-thinking mode) is available as open weights under the MIT license on Hugging Face and is the basis for self-hosting. On 31 July 2026 the API was updated to V4-Flash-0731 (per DeepSeek only re-post-trained; Terminal-Bench 2.1 82.7, NL2Repo 54.2, DeepSWE 54.4, DSBench-Hard 59.6); on 21 August 2026 the experimental V4-Flash-Vision-Exp followed (Terminal-Bench 2.1 83.9, NL2Repo 57.7, DeepSWE 59.3, DSBench-Hard 63.6, Chartography 64.3). Both builds have been superseded by V4.1-Flash since 10 September 2026. The weights of V4-Flash-0731 and V4-Flash-Vision-Exp are also available under the MIT license on Hugging Face.
Pricing
Since 16 August 2026 (16:00 UTC), DeepSeek bills by time of day. Peak time is Monday to Friday 01:00–04:00 and 06:00–10:00 UTC (CEST: 03:00–06:00 and 08:00–12:00); all other hours are off-peak with a 50% discount.
| Model | Cache hit | Cache miss (input) | Output |
|---|---|---|---|
| V4.1-Flash – peak | USD 0.006 | USD 0.30 | USD 1.20 |
| V4.1-Flash – off-peak | USD 0.003 | USD 0.15 | USD 0.60 |
| V4-Pro – peak | USD 0.044 | USD 1.32 | USD 3.96 |
| V4-Pro – off-peak | USD 0.022 | USD 0.66 | USD 1.98 |
All figures in USD per 1M tokens, prices per provider, as of October 2026. With the switch to V4.1-Flash, Flash prices have fallen compared with V4-Flash; per DeepSeek the 1017 build does not change prices. For batch and agent workloads, shifting to off-peak is worthwhile – from a European perspective that means afternoon, evening and night. V4-Pro costs roughly three times as much as V4.1-Flash.
Search and Fetch API: Since 17 October 2026 DeepSeek has also offered standalone Search and Fetch APIs built on its in-house search engine (web search and fetching web pages as Markdown). Per the pricing overview each costs USD 0.75 per 1,000 successful requests. Server-side tool calls via DeepSeek Harness, Claude Code or Codex hit the same APIs and are free until 1 November 2026 (00:00 Beijing time), after which they are billed per request. As with the chat API, DeepSeek processes requests on its platform in China.
Key strengths
- Open source and licensing: Weights are publicly available on GitHub and Hugging Face; R1 and the V4 and V4.1 repositories (including V4.1-Flash-1017 and V4-Pro-0813) are MIT-licensed, V3 is under a separate Model Agreement. Fine-tuning and customization are possible.
- MoE architecture: Mixture-of-Experts with only a fraction of parameters active (V4-Flash about 13B of 284B, V4-Pro about 49B of 1.6T; V3/R1: 37B of 671B), with Multihead Latent Attention and, in V4, hybrid attention (Compressed Sparse Attention + Heavily Compressed Attention).
- Reasoning (R1): DeepSeek-R1 makes chain-of-thought visible, is strong in mathematics and logic and is positioned as comparable to OpenAI o1.
- Languages: 50+ languages, best quality in English and Chinese, good quality in German.
EU availability and GDPR
Genuine EU data residency for DeepSeek exists at the three major hyperscalers only in a few configurations. New builds appear first on the direct DeepSeek API; the hyperscalers are at different stages: Google has offered DeepSeek-V4.1-Flash since 14 October 2026, Microsoft lists V4-Flash, V4-Flash-0731 and V4-Pro, and AWS still offers only V3.x and R1.
AWS Bedrock
- Regions (In-Region): Stockholm (eu-north-1) and London (eu-west-2) per the Bedrock model cards – Frankfurt (eu-central-1) and Ireland (eu-west-1) are not listed for DeepSeek. Stockholm is in the EU, London in the United Kingdom (outside the EU).
- Models: DeepSeek-V3.2 (
deepseek.v3.2, status Active) and V3.1 (deepseek.v3-v1:0ordeepseek.v3.1), each In-Region only, no geo or global routing. - Lifecycle: V3.1 has been Legacy since 30 September 2026, with end of life on 30 March 2027. Per AWS, Legacy models are no longer open to new customers; for new projects in the EU that leaves V3.2 in Stockholm.
- DeepSeek-R1: only in the US (us-east-1, us-east-2, us-west-2) via the “US” geo region, no EU region; also Legacy with end of life on 30 March 2027.
Microsoft Foundry (formerly Azure AI Foundry)
- Global Standard (worldwide processing, no data residency guarantee): DeepSeek-V3.2, V3.2-Speciale, V4-Flash, V4-Flash-0731 (preview) and V4-Pro. Resources can also be created in European regions, but per Microsoft inference can run in any Azure region where the model is deployed.
- Data Zone Standard EU (processing exclusively in EU member states): per the region availability table only DeepSeek-V4-Flash – not V4-Pro, V4-Flash-0731, V3.2 or V3.2-Speciale. Whether Microsoft will update to V4.1-Flash is not confirmed.
- Data Zone Provisioned EU (reserved capacity, processing in EU member states): per the region availability table DeepSeek-V3.2, incl. Germany West Central, France Central, Sweden Central and West Europe.
- Standard/Regional (strict single region): not available for DeepSeek models per Microsoft.
- Older entries for “V3” and “R1” as models sold directly by Azure no longer appear in the “Foundry Models sold by Azure” overview; check this if your deployment relies on them.
Google Cloud (Gemini Enterprise Agent Platform, formerly Vertex AI)
- DeepSeek-V4.1-Flash: available as a managed API (
deepseek-v4.1-flash-maas, launch stage Preview) since 14 October 2026; model availability and ML processing exclusively in the “global” region (worldwide processing). Google commits to the endpoint until at least 28 November 2026. - Shut-down endpoints: The MaaS endpoints for DeepSeek-V3.1, V3.2, R1 (0528) and DeepSeek-OCR were deprecated on 21 July 2026 and shut down on 21 October 2026; Google points to self-deployment via Model Garden as the alternative.
- No EU region: Neither Frankfurt (europe-west3) nor the Netherlands (europe-west4) is listed for DeepSeek.
Direct API and self-hosting
- Direct API: DeepSeek servers in China – caution with sensitive or personal data.
- Self-hosting: The open-source model can run in your own infrastructure: all data stays under your control, no dependency on external APIs, full GDPR compliance possible. This requires several high-end GPUs (A100/H100). Given the limited hyperscaler options, this is the most reliable path to genuine EU data residency.
Before any decision, check the specific model and the specific deployment type: “available in the EU” does not automatically mean that processing is limited to the EU.
Model lifecycle
- The API names
deepseek-chatanddeepseek-reasonerwere discontinued on 24 July 2026. The pricing overview listsdeepseek-flash(serving V4.1-Flash) anddeepseek-v4-pro. - V4-Flash-0731 and V4-Flash-Vision-Exp were withdrawn on 10 September 2026; their old API names are still accepted, requests are served by V4.1-Flash-1017 and billed at the Flash price.
- The V4.1-Flash-1017 build has replaced the 10 September V4.1-Flash build behind the API name
deepseek-flashsince 17 October 2026; the weights of both builds remain on Hugging Face. - V4-Flash and V4-Pro have been available in Microsoft Foundry since May 2026 (see above).
- On AWS Bedrock, DeepSeek-V3.1 and R1 have been Legacy since 30 September 2026, with end of life on 30 March 2027; V3.2 remains active.
- Google shut down the MaaS endpoints for V3.1, V3.2, R1 (0528) and DeepSeek-OCR on 21 October 2026 and has listed DeepSeek-V4.1-Flash as a preview since 14 October 2026.
- Peak/off-peak pricing has applied since 16 August 2026.
Our recommendation
DeepSeek is technically impressive and, per the vendor, delivers strong results in reasoning, coding and agent tasks. For cloud use, look closely: “usable in Europe” does not mean “data stays in the EU”.
- Cloud option with EU data residency: DeepSeek-V4-Flash via Microsoft Foundry with “Data Zone Standard EU”, or DeepSeek-V3.2 via AWS Bedrock in Stockholm (In-Region) or as “Data Zone Provisioned” EU in Microsoft Foundry. The direct API is an option for non-critical data; shifting batch workloads to off-peak halves token costs.
- Demanding agentic workflows: V4-Pro (GA build 0813) with reasoning effort
highormax; on Azure, however, only as “Global Standard” without an EU data zone. - Low-cost all-rounder with image understanding: V4.1-Flash-1017 via the direct API for non-critical data; per DeepSeek with shorter runtimes and outputs on agentic coding tasks.
- Self-hosting: DeepSeek-V4.1-Flash-1017 (MIT weights, 552B backbone parameters per the model card), V4-Flash or V3.2 for maximum control, customizability and the most reliable EU data residency; the V4-Pro-0813 weights are also published under MIT but require extreme amounts of hardware.
The choice depends on your requirements for control, compliance and technical resources – and with DeepSeek especially on the specific model and deployment type.