TL;DR
DeepSeek is a Chinese AI lab that releases large language models with open weights under the MIT license. As of September 2026, its main models are DeepSeek-V4-Pro and the newer, cheaper DeepSeek-V4.1-Flash. Both handle a 1 million token context and cost far less per token than GPT or Claude flagship models. Independent US government testing still puts DeepSeek behind the leading US models and flags weaker security against jailbreaks and agent hijacking. The hosted app and API store data in China, which is the main reason many companies and governments restrict them. Enterprises that want DeepSeek usually run the open weights on infrastructure they control, behind a gateway with logging, guardrails and evaluation.
Key Takeaways
- DeepSeek publishes open-weight models under the MIT license, so companies can download, fine-tune and run them commercially.
- The current API models are DeepSeek-V4.1-Flash (released September 10, 2026) and DeepSeek-V4-Pro, both with 1M token context.
- Official API rates run from $0.15 to $1.32 per million input tokens, with off-peak hours billed at half the peak rate.
- NIST’s CAISI found V4-Pro trails the US frontier by about eight months, yet it was more cost-efficient on 5 of 7 benchmarks.
- Risk depends on the access path, because the app and API process data in China while self-hosted weights keep data in-house.
- Production use needs a gateway, model routing, tool allow-lists and version-by-version evaluation before any output reaches users.
Watch on YouTube
DeepSeek: The AI Tool You Should Know About
Kanerika’s short explainer on what DeepSeek is, why its open models drew so much attention and what business teams should weigh before trying it.
Eight Months Behind, and Still the Price Everyone Benchmarks Against
In May 2026, the US Center for AI Standards and Innovation published its evaluation of DeepSeek-V4-Pro. Its verdict was blunt. However, the model trails the best US systems by roughly eight months, yet it beat the most cost-competitive US reference model on cost efficiency in 5 of 7 benchmarks, according to NIST’s published findings.
That tension is why this model family keeps showing up in enterprise AI budgets. As a result, a model that is good enough for a large share of workloads, at a fraction of the token price, forces every CIO to ask where it fits.
The harder part is the rest of the decision. In particular, who runs the model, where the data goes, how the outputs are checked and what happens when the next release lands three weeks later all matter more than the headline price.
What Is DeepSeek?
DeepSeek is an artificial intelligence company based in Hangzhou, China, founded in July 2023 by Liang Wenfeng and owned by the quantitative hedge fund High-Flyer, as its Wikipedia entry records. It also builds large language models and releases most of them with downloadable weights.
The company then became a global name in January 2025. Its R1 reasoning model launched on January 20 with performance DeepSeek described as on par with OpenAI o1, and the release notes put both code and weights under the MIT license. As a result, that license allowed commercial use, fine-tuning and distillation of the outputs.
Enterprises meet DeepSeek through three different surfaces, and each one carries a different cost and risk profile.
- The consumer app and website at chat.deepseek.com, which anyone can use for free because it needs no contract.
- The API platform, a paid pay-as-you-go service that also accepts both OpenAI and Anthropic request formats.
- The open weights on Hugging Face, which a company can run on its own GPUs or, alternatively, through a cloud provider.
Most confusion in news coverage comes from mixing these three up, especially in news stories about bans. Put simply, a ban on the phone app says nothing about a model running inside a private data center.
Open Weights, Not Open Data
The model files are open, but the training data is not. The DeepSeek-V4-Pro model card states that both the repository and the weights are licensed under MIT, while the pretraining corpus is described only in general terms.
That distinction matters for compliance teams because it changes what they can audit. In other words, you can inspect, host and modify the model, but you cannot audit exactly what went into it, so behavior testing has to replace data lineage review.
For a wider view of how these models sit among Llama, Qwen, Mistral and other open families, see Kanerika’s guide to open-source LLMs.
DeepSeek Models as of September 2026
The lab ships fast, and model names change with it. The table below therefore lists the releases that still matter for enterprise decisions, verified against the official API change log and the Hugging Face model cards on September 16, 2026.
| Model | Released | Size (total / active) | What Changed | Status in the API |
|---|
| DeepSeek-V3 | December 2024 | 671B / 37B | Mixture-of-experts general model | Superseded |
| DeepSeek-R1 | January 2025 | 671B / 37B | Reasoning model, MIT license, distilled small variants | Superseded |
| DeepSeek-V3.1 | August 2025 | 671B / 37B | Thinking and non-thinking modes in one model | Superseded |
| DeepSeek-V3.2 | December 2025 | 671B / 37B | DeepSeek Sparse Attention, stronger reasoning | Superseded |
| DeepSeek-V4-Pro | Preview April 24, 2026; GA August 13, 2026 | 1.6T / 49B | 1M context by default, hybrid compressed attention, stronger agents | Live as deepseek-v4-pro |
| DeepSeek-V4-Flash | Preview April 24, 2026; official July 31, 2026 | 284B / 13B | Smaller, faster V4 variant | Retired September 10, 2026, calls route to V4.1-Flash |
| DeepSeek-V4.1-Flash | September 10, 2026 | 552B backbone / 8B to 16B | New architecture, native image input, much smaller memory per token | Live as deepseek-flash |
Two practical points therefore stand out. The flagship choice today is between V4-Pro for the broadest knowledge and V4.1-Flash for speed, cost and image input. In addition, all current weights ship under MIT.
Even so, the lineup will keep moving, so treat this table as a snapshot and check the change log before you pin a version in production.
API Model Names Changed in 2026
Teams with older integrations should check their code first, because names have changed. The V4 release announcement retired the long-standing deepseek-chat and deepseek-reasoner names after July 24, 2026.
Instead, the current names are deepseek-flash for V4.1-Flash and deepseek-v4-pro for V4-Pro. The old deepseek-v4-flash name still works, but DeepSeek now routes it to V4.1-Flash and bills it at the Flash price, so a silent model change may already have happened in your pipelines.
Therefore, rather than hardcoding a model string in every application, route requests through one gateway configuration so a rename becomes a single change. Kanerika explains that pattern in its guide to the LLM gateway.
How DeepSeek Keeps Inference Cheap
The low prices come from architecture choices that cut how much hardware each token needs. In particular, three design choices do most of the work.
Mixture of experts. In other words, only a small slice of the model runs for each token. V4-Pro holds 1.6 trillion parameters but activates 49 billion per token, and V4.1-Flash activates 8 billion while reading input and 16 billion while writing output, according to the V4.1-Flash model card.
Compressed attention for long context. Normally, long prompts make memory costs explode, because every earlier token must stay in memory. V4 therefore combines Compressed Sparse Attention and Heavily Compressed Attention. As a result, at a 1M token context V4-Pro needs only 27% of the per-token compute and 10% of the memory cache that V3.2 needed.
Similarly, V4.1-Flash goes further with a second version of that design. Its memory cache takes about 890 bytes per token, roughly a quarter of V4-Flash, so million-token prompts become affordable.
Low-precision weights. Specifically, expert weights are stored in 4-bit precision and most other weights in 8-bit. As a result, that shrinks the GPU memory needed to host the model without retraining it at full precision.
These choices also set the self-hosting bill, although not always in the way buyers expect. A smaller active parameter count lowers compute per request. However, the full parameter set still has to sit in GPU memory, so V4-Pro remains a multi-GPU deployment even with low-precision weights.
In addition, the API exposes reasoning depth as a setting. Currently, V4-Pro and V4-Flash offer low, high and max thinking effort, while V4.1-Flash accepts a number from 1 to 100, so teams can trade answer quality for cost per call.
DeepSeek API Pricing in 2026
The API charges per million tokens, with cheaper rates for cached input and a 50% discount outside peak hours. Peak hours run from 1 a.m. to 4 a.m. and 6 a.m. to 10 a.m. UTC, Monday through Friday, according to the official Models and Pricing page.
| Price per 1M Tokens (as of Sept 16, 2026) | deepseek-flash (V4.1-Flash) | deepseek-v4-pro (V4-Pro) |
|---|
| Input, cache hit (off-peak / peak) | $0.003 / $0.006 | $0.022 / $0.044 |
| Input, cache miss (off-peak / peak) | $0.15 / $0.30 | $0.66 / $1.32 |
| Output (off-peak / peak) | $0.60 / $1.20 | $1.98 / $3.96 |
| Context / max output | 1M / 384K | 1M / 384K |
| Image input | Supported | Not supported |
| Concurrency limit | 2,500 | 500 |
For example, take a workload of 100 million input tokens and 20 million output tokens a month, with no cache hits. The comparison below then applies each vendor’s published standard rates from OpenAI’s pricing page and Anthropic’s pricing page.
| Model | Input / Output per 1M | Monthly Cost for 100M In + 20M Out |
|---|
| DeepSeek-V4.1-Flash (off-peak / peak) | $0.15 / $0.60 off-peak | $27 / $54 |
| OpenAI GPT-5.6 Luna | $0.20 / $1.20 | $44 |
| DeepSeek-V4-Pro (off-peak / peak) | $0.66 / $1.98 off-peak | $105.60 / $211.20 |
| Anthropic Claude Haiku 4.5 | $1 / $5 | $200 |
| Anthropic Claude Sonnet 5 | $2 / $10 | $400 |
| Anthropic Claude Opus 5 | $5 / $25 | $1,000 |
| OpenAI GPT-6 Astra | $10 / $50 | $2,000 |
The gap to frontier models is large. By contrast, the gap to small US models is not. Notice that GPT-5.6 Luna lands between the two DeepSeek tiers, which is a useful reminder that “cheapest” depends on the tier you are actually comparing.
Prices in this space change often. In fact, the pricing page itself notes that rates may change, so re-run the math against the live pricing pages before you commit a budget.
Token Price Is Only Part of the Cost
The per-token rate is the easiest number to compare, but it is also the least complete. Therefore, what a DeepSeek deployment costs depends mainly on how you access it.
| Cost Area | DeepSeek API | Managed Cloud Hosting | Self-Hosted Weights |
|---|
| Infrastructure | None, pay per token | Provider per-token or capacity pricing | GPU servers or reserved cloud GPUs |
| Data location | Processed in China | Your chosen cloud region | Your network |
| Engineering effort | Low | Low to medium | High, serving, scaling and patching |
| Model version control | Vendor can reroute names | Depends on provider catalog | Fully pinned by you |
| Governance and evidence | Hardest to evidence for audits | Provider controls plus yours | All controls are yours to build |
In addition, gateway, logging, evaluation and red-team testing belong in every column. However, those costs rarely appear in token comparisons, yet they usually decide whether a pilot reaches production. Document-heavy workloads such as intelligent document processing are a common first production use, because accuracy is easy to measure against a labeled sample.
Checklist
AI Governance Readiness Checklist
Check whether your policies, access controls, monitoring and model approval steps are ready before DeepSeek or any other new model reaches production.
Get the Checklist →How DeepSeek Compares With GPT and Claude on Benchmarks
The lab’s own numbers put its newest models close to US flagships. In particular, the V4.1-Flash model card compares it with Claude Opus 5 and GPT-5.6 Sol at maximum reasoning effort, and those figures are vendor-reported.
| Benchmark (reported by DeepSeek) | Claude Opus 5 | GPT-5.6 Sol | DeepSeek-V4-Pro | DeepSeek-V4.1-Flash |
|---|
| GPQA Diamond (science reasoning) | 93.4 | 94.1 | 92.4 | 90.9 |
| Humanity’s Last Exam, no tools | 56.3 | 44.5 | 42.7 (text-only subset) | 36.8 |
| Terminal-Bench 2.1 (agentic coding) | 89.1 | 88.8 | 87.9 | 90.6 |
| Terminal-Bench 4.0 (harder agentic tasks) | 51.8 | 39.9 | 12.4 | 31.2 |
| DeepSWE v1.1 (software fixes) | 74.0 | 73.0 | 62.7 | 74.2 |
Read the table by task type rather than by average, because the averages hide the gaps. On established tests the models sit close together. However, the hardest agentic benchmark still shows a wide gap to Claude Opus 5.
For side-by-side coverage of the US models, however, Kanerika’s Grok, ChatGPT and DeepSeek comparison, its comparison of ChatGPT, Gemini and Claude and its ranking of top LLMs all go deeper.
What Independent Testing Found
Independent evaluation, however, tells a more cautious story. For instance, CAISI, part of the US National Institute of Standards and Technology, ran V4-Pro on nine benchmarks, including two non-public tests that a model cannot have seen during training.
| CAISI Benchmark (April 2026) | GPT-5.5 | GPT-5.4 mini | Claude Opus 4.6 | DeepSeek-V4-Pro |
|---|
| SWE-Bench Verified | 81% | 73% | 79% | 74% |
| PortBench (non-public) | 78% | 41% | 60% | 44% |
| GPQA-Diamond | 96% | 87% | 91% | 90% |
| ARC-AGI-2 semi-private | 79% | Not reported | 63% | 46% |
CAISI concluded that the self-reported results placed V4 near Claude Opus 4.6 and GPT-5.4, while its own testing placed it closer to GPT-5, a model released about eight months earlier. At the same time, the same report found V4 more cost-efficient than GPT-5.4 mini on 5 of 7 benchmarks, ranging from 53% cheaper to 41% more expensive.
The lesson for buyers is therefore simple. Vendor benchmarks are only a starting point. Instead, your own evaluation set, built from real tasks and scored the same way for every model, is the only result that should drive a production choice. Product teams can build that evaluation set into their release process, as our guide to AI in product development describes.
Is DeepSeek Safe for Enterprise Use?
DeepSeek can be used safely in an enterprise, but not every way of using it is safe for sensitive data. Instead, the answer depends on the access path, the data involved and the controls around the model.
Start with the hosted service. First, DeepSeek’s privacy policy says it collects prompts, uploaded files and chat history, and that it directly collects, processes and stores personal data in the People’s Republic of China.
The service has also had a public security lapse. In January 2025, Wiz Research found an exposed DeepSeek database with more than a million log lines, including chat history and secret keys, which DeepSeek secured after disclosure.
Model behavior is the third concern, and it applies even when the weights are self-hosted. The 2025 CAISI evaluation found agents built on DeepSeek-R1-0528 were 12 times more likely than US frontier models to follow hijacking instructions. Furthermore, under a common jailbreak, that model answered 94% of overtly malicious requests, compared with 8% for US reference models, and CAISI also reported that DeepSeek models echo Chinese Communist Party narratives.
None of this rules the model out. Instead, it means the model needs the same guardrails you would put around any AI agent, with tighter limits on tools and data it can reach, as covered in Kanerika’s LLM security guide.
Watch on YouTube
Can LLM Gateways Make Enterprise AI Safer?
How a gateway layer adds authentication, logging and policy checks between applications and any model, including open-weight models like DeepSeek.
Why DeepSeek Is Restricted in Some Places
Restrictions so far target the hosted app and service, especially on government devices. So far, however, they have not banned the open-weight models themselves.
For private companies in the US, however, the service remains available. Therefore, the practical question is whether your own policies, customer contracts and regulators allow sending data to a service that stores it in China.
Risk Depends on How You Access It
Splitting the platform into its access paths therefore turns a vague safety debate into a concrete review. That is because each path answers the data question differently.
| Access Path | Where Prompts Go | Main Risk | Suitable For |
|---|
| Consumer app or website | DeepSeek servers in China | Staff pasting confidential data | Public information only, often blocked by policy |
| DeepSeek API | DeepSeek servers in China | Data residency and contract gaps | Non-sensitive, public or synthetic data |
| Managed cloud hosting | Your cloud provider’s region | Model behavior, provider catalog lag | Internal workloads under existing cloud agreements |
| Self-hosted open weights | Your own network | Jailbreaks, agent hijacking, operating cost | Sensitive data with strong guardrails |
Across these paths, model behavior risk stays constant, while data risk falls as control moves in-house. Consequently, regulated companies considering DeepSeek should start their evaluation with self-hosted or cloud-hosted weights.
Three Ways to Run DeepSeek in the Enterprise
Enterprises typically reach DeepSeek through the API, a managed cloud model catalog or their own GPUs. Each option moves more control, and more work, to your team. The infographic also shows the consumer chat app as a baseline for comparison.
1. The DeepSeek API. This is the fastest way to test, so most teams begin here. The API uses the base URL api.deepseek.com and also accepts OpenAI Chat Completions, the OpenAI Responses format and the Anthropic Messages format, so most existing SDKs work with a changed model name.
2. Managed cloud hosting. Major clouds also offer DeepSeek models inside their own AI platforms. Amazon Bedrock lists DeepSeek models as fully managed, serverless options, which keeps prompts inside your AWS account boundary, although catalogs often trail DeepSeek’s newest release.
3. Self-hosted open weights. Here, teams download the weights from Hugging Face and then serve them with an inference engine such as vLLM or SGLang. As a result, you get full control over versions, data and network access, but you also take on the GPU bill and operations.
Self-hosting also means choosing the right engine for your scale. For guidance, Kanerika’s comparison of vLLM and Ollama covers when a lightweight runner is enough and when a production server is needed, and its overview of private LLMs explains the wider operating model.
What to Have in Place Before Self-Hosting
Running a frontier-size open model is an infrastructure project rather than a download. Consequently, these are the items teams most often underestimate.
- GPU capacity sized to the full model. V4-Pro stores 1.6 trillion parameters, and CAISI served V4 on cloud H200 and B200 GPUs for its tests, so plan for a multi-GPU cluster rather than a single card.
- A pinned model version. Record the exact Hugging Face revision you tested, and then never swap versions without rerunning your evaluation set.
- An evaluation suite. Build a fixed set of real tasks, including prompts designed to jailbreak or hijack the model, and score every candidate model the same way.
- Network isolation for agents. Give the model only the tools and data sources a task needs, and also keep outbound internet blocked by default.
- A license and policy review. MIT allows commercial use. Even so, legal and security teams should still sign off on using a PRC-developed model for each use case.
In practice, teams that treat these steps as a checklist get to a stable pilot much faster than teams that start with the GPUs and work backward.
How to Govern DeepSeek in Production
A governed deployment puts a control layer between users and the model, applying the same principles as any AI governance program. The layer then decides which model answers, what data it can see, what it is allowed to do and how every step is recorded.
In practice, that layer therefore has five parts.
- An LLM gateway. Every application calls one endpoint, which then handles authentication, rate limits, cost tracking and model names in one place.
- Model routing rules. Low-risk, high-volume tasks go to a cheaper model such as V4.1-Flash, while sensitive or complex tasks go to a model that passed stricter tests.
- Input and output guardrails. Personal data is redacted before prompts leave the application, and responses are checked for policy violations before users see them.
- Tool allow-lists for agents. Because of CAISI’s hijacking results, agents built on DeepSeek should only call approved tools with scoped credentials and human approval for high-impact actions, whether those tools connect directly or through the Model Context Protocol.
- Evaluation gates and audit logs. Each new model version must pass the same test set before routing changes. Also, prompts, outputs and tool calls are logged for review, which is the core of LLMOps observability.
This design also keeps you portable, because no application depends on one model. When a new release underperforms, or when a regulator changes its position, the gateway lets you shift traffic to another model without rewriting applications.
However, measuring quality over time is where many programs stall, so Kanerika’s LLM evaluation framework lays out how to build scoring that holds up across model changes.
Where DeepSeek Fits and Where It Doesn’t
DeepSeek fits best where volume is high, data can be controlled and outputs are checked before they matter. By contrast, it fits worst when an error is expensive and the model acts on its own.
| Workload | Fit | Recommended Access Path | Why |
|---|
| Summarizing and classifying large document sets | Strong | Self-hosted or cloud-hosted | Long context and low cost per token |
| Internal knowledge search with retrieval | Strong | Self-hosted behind access controls | Answers stay grounded in approved sources |
| Code generation with human review | Good | Any path, with no secrets in prompts | Competitive coding scores, reviewer catches errors |
| Customer-facing chat | Limited | Self-hosted with strict output filters | Jailbreak and narrative risks reach the public |
| Autonomous agents with system access | Weak today | Only in sandboxes with approvals | Hijacking results and lower scores on hard agentic tests |
| Regulated personal or financial data via hosted API | Poor | Avoid the hosted API | Data stored in China |
As a result, many teams end up with a mixed setup. DeepSeek handles bulk summarization and retrieval-augmented generation behind the gateway, while a US frontier model handles agentic and customer-facing tasks, which is also the pattern Kanerika describes for small language models in cost-sensitive stacks.
Kanerika Service
LLM Development and Model Routing
Kanerika designs and deploys LLM applications that route each task to the right model, open-weight or hosted, with evaluation, guardrails and cost tracking built in.
Explore LLM DevelopmentDeepSeek Limitations to Plan Around
The weaknesses are specific, so they stay manageable when you plan for them early.
- A real capability gap on hard tasks. Independent tests still place V4-Pro months behind the US frontier, with the widest gaps on cyber, abstract reasoning and non-public software tasks.
- Censorship and narrative bias. For example, CAISI found DeepSeek models repeat Chinese Communist Party positions, which matters for research, policy and public-facing content.
- Weaker resistance to jailbreaks and hijacking. As a result, this raises the bar for guardrails on any agent or chatbot built on the model.
- No published enterprise plan. The official pricing page lists pay-as-you-go rates and concurrency limits but no enterprise tier or service-level terms.
- Disputed training practices. In a February 2026 memo to the US House Select Committee on China, OpenAI said it had observed DeepSeek-linked accounts trying to distill its models, as Rest of World reported. That allegation is a reason to involve legal and procurement teams early.
- Uneven features across models. Image input works on V4.1-Flash but not V4-Pro. Similarly, concurrency limits differ five-fold between them.
- Fast release churn. In total, five model releases or retirements reached the API between April and September 2026, so version pinning and regression testing are not optional.
Finally, hallucination remains a risk with every language model. Kanerika’s guide to LLM hallucination covers the grounding and verification techniques that reduce it.
Common Mistakes When Adopting DeepSeek
The same few mistakes show up again and again in enterprise LLM pilots, and this model family makes several of them more costly.
- Choosing on price alone. A cheap model that fails your hardest tasks actually costs more when rework and review time are counted.
- Trusting vendor benchmarks. Self-reported and independent scores diverge, so test on your own data first.
- Letting staff use the public app. Without a clear policy and blocking, however, confidential data ends up in a service that stores it abroad.
- Giving agents broad tool access. Hijacking risk turns a helpful agent into an attack path when it can reach email, files or payments.
- Skipping regression tests on upgrades. A renamed or rerouted model can change behavior overnight, even when your own code has not changed.
Overall, each of these is cheap to prevent before launch but expensive to fix after users depend on the system.
How to Decide If DeepSeek Belongs in Your Stack
A short, structured review keeps the decision grounded. Then work through these five questions with IT, security, legal and the business owner in the room, ideally as part of a wider AI strategy review.
| Question | What a Good Answer Looks Like | If the Answer Is Weak |
|---|
| What task are we solving? | High-volume, reviewable work with a clear quality bar | Define the use case before comparing models |
| What data will the model see? | Classified data with a matching access path | Rule out the hosted API for sensitive data |
| Where will it run? | Budgeted GPUs or an approved cloud catalog | Start with a cloud-hosted pilot |
| How will we measure quality? | A fixed evaluation set scored for every model | Build the test set before the pilot |
| Who owns governance? | A named owner, a gateway and audit logs | Hold production launch until this exists |
In short, if three or more answers are weak, the organization is not ready for any new model in production, DeepSeek or otherwise. For that reason, fixing readiness first pays off across every AI project that follows.
How Kanerika Helps Enterprises Run Open-Weight and Hosted Models Safely
Kanerika is a data and AI consulting firm that builds AI agents, LLM applications and AI governance programs for enterprises. Kanerika builds enterprise solutions on both Anthropic’s Claude and OpenAI’s models, and is an OpenAI Select Partner, so model choice follows the use case rather than a single vendor contract.
For example, for clients weighing any open-weight model, Kanerika works in five stages.
- Assess. First, classify the target workloads and data, and then map each to an allowed access path.
- Evaluate. Next, build a task-based test set, including jailbreak and hijacking prompts, and score candidate models side by side.
- Architect. Then stand up the gateway, routing rules, guardrails and hosting, whether self-hosted, cloud-hosted or hybrid.
- Govern. After that, put access controls inside retrieval, log every prompt and tool call, and assign an owner for model changes.
- Operate. Finally, monitor cost and quality, and rerun evaluations before any model version or route changes.
The governance stage also draws on work Kanerika has already delivered in regulated settings. For example, at a leading investment bank, Kanerika deployed its KlarityIQ agent for unstructured documents and its Karl agent for database queries, with role-based access control enforced at the retrieval layer.
The published results were 43% faster information retrieval, 100% role-based compliance at the point of retrieval and 35% higher workforce efficiency, according to the investment bank case study. That engagement did not use this model. Even so, it shows the control that matters most when any model touches sensitive data, which is access enforced inside retrieval rather than bolted on afterward.
Case Study
43% Faster Retrieval for an Investment Bank
Kanerika deployed KlarityIQ and Karl with role-based access enforced at the retrieval layer, achieving 43% faster information retrieval, 100% role-based compliance and 35% higher workforce efficiency.
Read the Case Study →Wrapping Up
DeepSeek offers open-weight models that come close to US flagships on many tests, while charging a fraction of the token price. However, independent testing still shows a capability gap, weaker jailbreak resistance and a hosted service that stores data in China.
Ultimately, the deciding factor is how you run it. Self-hosted or cloud-hosted weights behind a gateway, with routing, guardrails and evaluation, therefore turn DeepSeek into a useful, controllable part of an enterprise AI stack.
Frequently Asked Questions
What is DeepSeek?
DeepSeek is a Hangzhou-based AI company, founded in 2023 and owned by the hedge fund High-Flyer, that builds large language models. It releases model weights under the MIT license and also runs a free chat app and a paid API. As of September 2026, its main models are V4-Pro and V4.1-Flash.
Is DeepSeek free to use?
The DeepSeek chat app and website are free for individual use. The API is paid per million tokens, starting at $0.15 for input and $0.60 for output on V4.1-Flash during off-peak hours as of September 2026. The open weights are free to download, but running them requires your own GPU infrastructure or a cloud provider.
Is DeepSeek safe to use for business data?
It depends on how you use it. The company’s privacy policy says the app and API store data in China, so sensitive data should not go there. Self-hosted or cloud-hosted weights keep data under your control, but US government testing found weaker jailbreak and agent-hijacking resistance, so strong guardrails are still required.
Why is DeepSeek banned in some places?
Restrictions focus on the hosted app and data transfers to China. Italy’s data protection authority limited processing of Italian users’ data in January 2025, Texas banned the app on state-issued devices, and a US House bill proposes removing it from federal devices. These rules target government use and the service, not the open-weight models.
Is DeepSeek available in the US?
Yes. As of September 2026, US companies and individuals can use the app, the API and the open weights. Some state governments and federal agencies restrict it on official devices. Private companies should check their own security policies, customer contracts and regulators before sending any data to the hosted service.
Is DeepSeek better than ChatGPT?
It is much cheaper per token and its weights can be self-hosted, which ChatGPT’s models cannot. On quality, independent NIST testing placed V4-Pro about eight months behind the best US models, with larger gaps on hard agentic and cyber tasks. For many bulk text tasks it is good enough, so test both on your own workload.
Is DeepSeek open source?
It is open weight. Its model files and code are published on Hugging Face under the MIT license, which allows commercial use, modification and fine-tuning. The training data is not released, so you can inspect and run the model but cannot fully audit what it learned from.
What hardware do you need to self-host DeepSeek?
The flagship models need data-center GPUs. V4-Pro has 1.6 trillion parameters, and NIST served V4 on H200 and B200 GPUs during testing, so plan for a multi-GPU cluster. V4.1-Flash is smaller at a 552 billion parameter backbone but still needs server-class GPUs and an inference engine such as vLLM or SGLang.
Which DeepSeek model should enterprises start with?
Start with V4.1-Flash for high-volume summarization, classification, retrieval and coding help, because it is cheaper, faster and accepts images. Use V4-Pro where broader knowledge matters more than cost. For either model, run your own evaluation set first and keep a US frontier model available for hard agentic tasks.
What changed between DeepSeek V4 and V4.1?
V4.1-Flash, released September 10, 2026, replaced V4-Flash in the API. It uses a new architecture that activates 8 billion parameters while reading input and 16 billion while writing, cuts memory per token to about a quarter of V4-Flash, and adds native image input. V4-Pro remains available unchanged.
