DeepSeek R1: Current Status, R1-0528, and What to Use Today

Last verified: October 4, 2026

DeepSeek R1 was one of DeepSeek’s most important reasoning model releases. It made DeepSeek widely known outside China, introduced many developers to DeepSeek’s reasoning-style outputs, and remains available as an open-weight model family. But there is an important catch: DeepSeek R1 should not be treated as the current default DeepSeek API model for new hosted API projects unless DeepSeek’s current API docs explicitly say so.

The short answer: DeepSeek R1 is still relevant as an open-weight reasoning model family and historical milestone. The later DeepSeek-R1-0528 update is also still important for people who specifically want the R1 family. But old tutorials that tell you to call deepseek-reasoner as the default hosted API model may be outdated. DeepSeek’s changelog later moved the API through V3.1, V3.2, V4, and V4.1, and the current hosted API direction uses newer model names such as deepseek-flash and deepseek-v4-pro. DeepSeek API Change Log

If you want R1 specifically, use the official open weights or a compatible local/third-party inference stack. If you want hosted DeepSeek reasoning through the official API, check the current Models & Pricing page and use the currently documented model IDs and Thinking Mode controls.

Quick answer: is DeepSeek R1 still available?

Yes, but the answer depends on what you mean by “available.”

• As open weights: Yes. The official DeepSeek-R1 Hugging Face model card is still live and lists DeepSeek-R1 and DeepSeek-R1-Zero as 671B-total-parameter, 37B-activated-parameter models with 128K context length. DeepSeek-R1 on Hugging Face
• As R1-0528: Yes. DeepSeek-R1-0528 is a later R1 update with its own official model card and usage notes. DeepSeek-R1-0528 on Hugging Face
• As deepseek-reasoner in new API code: Be careful. The changelog shows deepseek-reasoner was introduced for R1 in January 2025, upgraded to R1-0528 in May 2025, later mapped to newer model generations, and then scheduled for discontinuation as a legacy API model name in 2026.
• As the current hosted DeepSeek API default: No source checked supports treating R1 as the current default hosted API model. For new API work, use the current model IDs listed in DeepSeek’s active API documentation.

So the practical answer is: use R1 or R1-0528 if you specifically want the R1 open-weight reasoning family; use newer DeepSeek API models if you want current hosted API access.

DeepSeek R1 status table

The table below separates model weights, hosted API aliases, and current DeepSeek API model names. This distinction is the most important part of using DeepSeek R1 correctly today.

What DeepSeek R1 was and why it mattered

DeepSeek R1 was DeepSeek’s first major reasoning model family. It became important because it combined strong reasoning behavior with open-weight availability and permissive licensing, making it unusually visible to developers, researchers, and AI builders.

The official DeepSeek-R1 model card lists two main full-size models: DeepSeek-R1-Zero and DeepSeek-R1. Both are listed with 671B total parameters, 37B activated parameters, and 128K context length. The same card says they are trained based on DeepSeek-V3-Base. DeepSeek-R1 model card

R1 also came with distilled model releases based on Qwen and Llama models. Those distill models made the R1 family more practical for local experimentation because the full 671B model is not realistic for many individual developers or small teams to run directly.

For searchers today, the key point is that R1 is not just a historical name. The model card and weights still matter. But the hosted DeepSeek API has moved forward, so R1’s role has changed from “the API model everyone should call” to “an important open-weight reasoning family and reference point.”

DeepSeek R1 vs DeepSeek-R1-0528

DeepSeek-R1-0528 is not the same thing as the original R1 release. It is a later R1 update. DeepSeek’s changelog says deepseek-reasoner was upgraded to DeepSeek-R1-0528 on May 28, 2025, with stronger reasoning, improved front-end generation, reduced hallucinations, and added JSON Output and Function Calling support. DeepSeek API Change Log

The official R1-0528 model card also includes updated local-use guidance. It says system prompts are now supported and that it is no longer required to add <think>\n at the beginning of the output to force the model into a thinking pattern. DeepSeek-R1-0528 model card

If you are starting a new local R1 experiment today, R1-0528 is usually the more relevant R1-family model to check first. If you are studying the original release or reproducing older results, use the original R1 model card and paper context.

What happened to deepseek-reasoner?

deepseek-reasoner is the source of much of the confusion around DeepSeek R1. It was an API model name, not the same thing as the permanent identity of the R1 open-weight model family.

DeepSeek’s changelog shows this lifecycle:

• January 20, 2025: deepseek-reasoner was announced as the new DeepSeek-R1 model name in the API.
• May 28, 2025: deepseek-reasoner was upgraded to DeepSeek-R1-0528.
• August–December 2025: the changelog shows deepseek-reasoner being upgraded along with newer DeepSeek-V3.1 and V3.2 thinking modes.
• April 24, 2026: DeepSeek’s V4 changelog says deepseek-chat and deepseek-reasoner would be discontinued after three months, with those names temporarily pointing to non-thinking and thinking modes of deepseek-v4-flash.
• September 10, 2026: DeepSeek’s V4.1 Flash entry says V4.1 Flash is available through deepseek-flash, while retired V4 Flash aliases route temporarily to V4.1 Flash.

The practical conclusion is simple: do not build new hosted API code around deepseek-reasoner unless the current API documentation at the time of your deployment explicitly tells you to. For current hosted API work, use the current model IDs shown in DeepSeek’s Models & Pricing and changelog documentation.

Does DeepSeek R1 still work in the API?

The safest answer is: do not assume it does.

The R1-0528 Hugging Face model card includes a section saying users can chat with DeepSeek-R1 on the official website by switching on “DeepThink,” and that DeepSeek provides an OpenAI-compatible API at its platform. However, API availability and model names changed after R1-0528. The DeepSeek changelog later shows deepseek-reasoner moving through newer model generations and then being scheduled for discontinuation as a legacy API name. DeepSeek-R1-0528 model card

That means an old tutorial from 2025 using model="deepseek-reasoner" may not be the right guide for a new 2026 API integration. It may fail, be redirected, or behave differently from the original R1 model depending on DeepSeek’s current routing rules.

For hosted API usage, treat DeepSeek’s current API docs as the source of truth. At the time this article was verified, current API work should be checked against model names such as deepseek-flash and deepseek-v4-pro, plus the current Thinking Mode documentation.

What should you use instead of R1 today?

The right replacement depends on what you wanted from R1 in the first place.

Do not make this decision from the model name alone. Choose based on hosting, latency, cost, context length, tool support, privacy, license, and operational requirements.

When DeepSeek R1 still makes sense

DeepSeek R1 still makes sense in several cases:

• You are researching reasoning-model behavior and want the original R1 family.
• You want open weights under the R1 model-card license terms.
• You want to compare old R1 results with newer DeepSeek models.
• You are experimenting with local inference or third-party inference providers.
• You want to test R1 distill models on smaller hardware or in local development.
• You are studying distillation, reinforcement learning, or reasoning traces.

R1 is not useless because it is no longer the default hosted API path. It is still important as an open-weight reasoning family. The key is to use it in the right context.

When to use DeepSeek V4.1 Flash or Thinking Mode instead

Use newer DeepSeek API models when you want hosted API access with current DeepSeek support. DeepSeek’s V4.1 Flash release says the model is live on the API with native multimodal support and that users should set the model to deepseek-flash. The same release says older V4 Flash names are retired and temporarily route to V4.1 Flash for compatibility. DeepSeek V4.1 Flash release note

Use Thinking Mode when you want reasoning behavior in the current API. DeepSeek’s Thinking Mode guide documents how to enable or disable thinking, use reasoning_effort, and handle reasoning_content. DeepSeek Thinking Mode guide

The practical migration is not “R1 becomes V4.1 Flash.” It is more accurate to say: R1 was the earlier reasoning model family; current hosted API reasoning now belongs to newer DeepSeek API models and Thinking Mode controls.

Open weights, license, and local use

DeepSeek-R1’s official model card says the code repository and model weights are licensed under the MIT License. It also says the DeepSeek-R1 series supports commercial use and allows modifications and derivative works, including distillation for training other large language models. DeepSeek-R1 on Hugging Face

The same card adds important detail for distill models: the Qwen-based distill models are derived from Qwen 2.5 models originally licensed under Apache 2.0, while the Llama-based distill models are derived from Llama 3.1 or Llama 3.3 models and are subject to their original Llama licenses.

Use precise language here. DeepSeek R1 is often called “open source” casually, but the more precise phrase is open weights. The model weights are available, but that does not mean self-hosting is cheap or simple.

The full R1 model is large. For many local experiments, distilled models are more realistic. The official R1 model card shows a vLLM example for a distilled Qwen model, and the R1-0528 card points users to the DeepSeek-R1 GitHub repository for local-running details. Do not assume full DeepSeek-R1 will run well on a normal laptop unless you are using a smaller distilled or quantized variant designed for that environment.

R1 distill models explained

DeepSeek-R1-Distill models are smaller models fine-tuned on samples generated by DeepSeek-R1. They are not the same as the full R1 model, but they can be more practical for local or lower-cost experimentation.

The original model card lists several distill models:

• DeepSeek-R1-Distill-Qwen-1.5B
• DeepSeek-R1-Distill-Qwen-7B
• DeepSeek-R1-Distill-Llama-8B
• DeepSeek-R1-Distill-Qwen-14B
• DeepSeek-R1-Distill-Qwen-32B
• DeepSeek-R1-Distill-Llama-70B

The R1-0528 model card also describes DeepSeek-R1-0528-Qwen3-8B, a smaller model produced by distilling chain-of-thought from DeepSeek-R1-0528 into Qwen3 8B Base. The card says its architecture is identical to Qwen3-8B but it shares the tokenizer configuration of DeepSeek-R1-0528. DeepSeek-R1-0528 model card

Benchmarks and historical importance

The DeepSeek-R1 and DeepSeek-R1-0528 model cards contain first-party benchmark tables. Those tables are useful for understanding why R1 mattered and what changed in R1-0528, but they should not be treated as proof that R1 is better than every current 2026 DeepSeek model.

The original R1 model card reports model-card evaluation results across categories such as English knowledge, math, code, and reasoning-style tasks. It also explains that, for all listed models, the maximum generation length was set to 32,768 tokens in those evaluations, and for sampling benchmarks the team used temperature 0.6 and top-p 0.95. DeepSeek-R1 model card

The R1-0528 model card shows improved results across several benchmark categories and notes updated evaluation settings, including a maximum generation length of 64K tokens for the listed model-card evaluations. These are first-party model-card numbers, not a substitute for testing your own workload. DeepSeek-R1-0528 model card

If you are choosing a model today, run your own evaluation. Test the exact model, prompt style, token budget, latency, output format, tool stack, and data sensitivity you actually need.

Common mistakes from old DeepSeek R1 tutorials

Practical checklist: what should you use today?

• Use deepseek-flash if you want current hosted DeepSeek API access and V4.1 Flash fits your task.
• Use Thinking Mode if you want reasoning behavior through current DeepSeek API models.
• Use DeepSeek-R1-0528 if you specifically want the updated R1-family open-weight model.
• Use original DeepSeek-R1 if you are reproducing, studying, or comparing the original R1 release.
• Use R1 distill models if you want smaller local reasoning experiments.
• Do not use deepseek-reasoner in new code unless the current API docs at your deployment time explicitly support it.
• Do not assume old R1 prompting advice still applies; check R1-0528 and current Thinking Mode docs.
• Do not assume open weights mean easy self-hosting; check infrastructure, inference stack, and license requirements.

Frequently asked questions

What is DeepSeek R1?

DeepSeek R1 is DeepSeek’s first major reasoning model family. It includes DeepSeek-R1-Zero, DeepSeek-R1, and distilled variants based on Qwen and Llama families. It is important historically and remains available as open weights.

Is DeepSeek R1 still available?

Yes, as open weights. The official DeepSeek-R1 and DeepSeek-R1-0528 Hugging Face model cards are still available. That is different from saying R1 is the current default hosted API model.

Is DeepSeek R1 still working?

As open weights, yes. For hosted API usage, do not assume old R1 API examples still work. Check DeepSeek’s current Models & Pricing page and changelog before using old model names such as deepseek-reasoner.

Is DeepSeek R1 the same as deepseek-reasoner?

No. DeepSeek R1 is a model family. deepseek-reasoner was an API model name that originally pointed to R1, was later upgraded to R1-0528, then mapped to newer thinking-mode models, and was scheduled for discontinuation as a legacy alias.

Is deepseek-reasoner deprecated?

DeepSeek’s April 2026 changelog says deepseek-chat and deepseek-reasoner would be discontinued after three months. Treat it as a legacy API name unless current DeepSeek docs explicitly say otherwise.

What is DeepSeek-R1-0528?

DeepSeek-R1-0528 is a later R1 update from May 2025. DeepSeek’s changelog says deepseek-reasoner was upgraded to R1-0528, and the official model card includes updated usage recommendations and benchmark tables.

Is DeepSeek-R1-0528 better than R1?

DeepSeek’s own changelog and model card describe R1-0528 as an upgrade with stronger reasoning, lower hallucinations, and improved features. Treat those as first-party claims and test your own workload before making a production decision.

Can I use DeepSeek R1 through the API?

Do not assume the original R1 is available through the current hosted API. Old deepseek-reasoner examples may be outdated. Use current DeepSeek API documentation as the source of truth for hosted model names.

What replaced DeepSeek R1 in the API?

DeepSeek’s API moved through later generations, including V3.1, V3.2, V4, and V4.1. For current hosted API use, check model names such as deepseek-flash and deepseek-v4-pro, plus the current Thinking Mode documentation.

Should I use DeepSeek R1 or DeepSeek V4.1 Flash?

Use DeepSeek V4.1 Flash or another current API model for hosted API projects. Use R1 or R1-0528 if you specifically need the open-weight R1 family for research, local inference, or model comparison.

Is DeepSeek R1 open source?

The more precise phrase is open weights. The official model card says the code repository and model weights are under the MIT License and support commercial use, modifications, derivative works, and distillation. Distilled models may also involve Qwen or Llama base-model license considerations.

Can I run DeepSeek R1 locally?

Yes, but full DeepSeek R1 is very large. Many users will find distilled or quantized variants more practical for local experimentation. Follow official model-card and repository guidance, and do not assume full R1 will run well on consumer hardware.

Do I still need to add <think>\n?

For DeepSeek-R1-0528, the official model card says it is not required to add <think>\n at the beginning of the output to force the model into a thinking pattern.

Should I use R1 for production?

Use R1 in production only if you have a clear reason to use the R1 open-weight family and you can support the operational requirements. For most new hosted API projects, start with current DeepSeek API model names and Thinking Mode controls.

Bottom line

DeepSeek R1 is still important, but its role has changed. It remains valuable as an open-weight reasoning model family and research milestone. DeepSeek-R1-0528 is the later R1 update and is the better R1-family reference for many newer local or research workflows.

For new hosted API code, do not blindly copy old deepseek-reasoner examples. Check the current DeepSeek API changelog, use currently documented model IDs such as deepseek-flash where appropriate, and use Thinking Mode when you want reasoning behavior in current DeepSeek API models.

Official sources checked

• DeepSeek API Change Log
• DeepSeek-R1 model card on Hugging Face
• DeepSeek-R1-0528 model card on Hugging Face
• DeepSeek V4.1 Flash release note
• DeepSeek Thinking Mode guide