DeepSeek R1 0528
Free
DeepSeek R1 0528 is an upgraded version of the inference model released by DeepSeek on 2025-05-28. The official model card clearly emphasizes its significant enhancements in mathematics, programming and general logic, and continues to support the web-side DeepThink experience and OpenAI compatible API calls.
DeepSeekR1 0528
Core parameters and statistics
In a word, DeepSeek R1 0528 is not a minor repair of "changing the date suffix", but a practical upgrade that advances the reasoning line of R1 from "can solve difficult problems" to "more suitable as a production model". The model card openly states that it significantly enhances the performance of mathematics, programming and general logic tasks through deeper thinking processes and post-training algorithm optimization, while simultaneously pushing function call support, code experience and illusion control a step forward.
| Projects | Public Information |
|---|---|
| Product form | Open source inference model + web version DeepThink + OpenAI compatible API |
| Latest version | DeepSeek-R1-0528 |
| Release time | 2025-05-28 |
| License | MIT |
| Model size | Hugging Face displays about 685B params |
| Maximum generated length | 64K tokens |
| Online portal | chat.deepseek.com, switchable DeepThink |
| API base address | https://api.deepseek.com |
| Corresponding web capabilities | File upload, Internet search, in-depth thinking |
| Contextual Main Line | API documentation currently sells the V4 series for 1M, and the previous generation R1-0528 focuses on reasoning depth rather than ultra-long contextual marketing |
| Community popularity | Hugging Face has about 2.46k likes and about 1.41 million downloads in the last 30 days |
Core positioning: It is not an all-round "chat skin upgrade", but a model version that is more like a "reasoning engine", suitable for disassembling complex problems, completing the mathematical process, and completing the coding steps.
Propaganda Verification: The official emphasizes “stronger reasoning depth”. Judging from the benchmark changes disclosed in the model card, this selling point is not an empty slogan. Especially in tasks such as AIME 2025, LiveCodeBench, Aider Polyglot, and SWE Verified that require continuous deduction, the improvement is clearly visible.
User and market recognition
The market acceptance of DeepSeek R1 0528 is more like "developers and heavy users vote first" than the large-scale seat procurement story of traditional SaaS. The reason is simple: what really impresses people is not the UI, but the open source inference model that for the first time brings difficult mathematics, complex code and multi-step logic to a level very close to that of the closed-source model.
A brief comment: If the team used to regard the closed-source inference model as a "guaranteed machine for difficult tasks", then the significance of R1 0528 is to reduce this part of the budget, not to replace all daily dialogue scenarios.
Publicity verification: The model card directly benchmarks the overall performance against O3 and Gemini 2.5 Pro. The more verifiable parts are in the sub-benchmarks: AIME 2025 improved from the previous version 70.0 to 87.5, AIME 2024 reached 91.4, LiveCodeBench improved from 63.5 to 73.3, Aider Polyglot improved from 53.3 to 71.6, and SWE Verified improved from 49.2 to 57.6. These numbers show that it has indeed captured the pain point that target users are most concerned about, which is the stability of difficult reasoning tasks.
Market Signal: Retaining the DeepThink switch on the web page means that DeepSeek does not lock such capabilities in papers and model cards, but continues to allow ordinary users to directly perceive the "slower but stronger" reasoning form. For productization, this perceived difference is more important than pure benchmarking.
Boundary reminder: It does not have a comprehensive performance on indicators such as SimpleQA that emphasize the accuracy of concise facts, which also explains why it is more suitable for "problem solving" instead of all Q&A being cut into the deepest thinking mode by default.
Cost advantage
The greatest commercial significance of DeepSeek R1 0528 is not the "free chat" itself, but the ability to move complex reasoning from a capability reserved only for high-budget teams to a range that ordinary developers and small and medium-sized teams dare to use regularly.
The truth about free: The web and app have indeed lowered the trial threshold, but free does not mean unlimited enterprise availability. When it comes to batch code review, automatic question answering, and long-process reasoning, costs and concurrency constraints will still fall on the API platform.
| Cost Hierarchy | Disclosure | Real Meaning |
|---|---|---|
| C-side/Personal | The chat portal can directly experience DeepThink | Suitable for manual use, not suitable for systematic calls |
| API/Developer | The current document mainly recommends V4, with prices starting from $0.0028 per million input cache hits, $0.14 per million misses, and $0.28 per million outputs | Description DeepSeek continues to maintain a low-price strategy, and the inference cost is much lower than most closed-source head models |
| Enterprise | Need to confirm based on actual concurrency, quota, and isolation strategy | The real big deal is concurrency, stability, and integrated governance, not just token unit price |
Implicit cost: The deeper the reasoning, the higher the output token consumption. The model card mentions that the average number of thinking tokens per question in the AIME test increased from 12K to 23K, which directly shows that "stronger" is often accompanied by higher reasoning overhead. Individual users are not very perceptive, but enterprise batch scenarios will be very obvious.
Compliance and Risk: The MIT license allows more room for self-deployment and distillation of models, but when integrating models into high-risk business processes, compliance does not mean that they are worry-free by default just because they are “open source”. When it comes to financial conclusions, medical advice, legal text or automated code submissions, a human confirmation point must still be added.
Hidden benefits: If the team originally needs to "run the cheap model first and review the expensive model once," inference models such as R1 0528 will reduce the number of reviews, and the rework rate and model switching costs will decrease.
Main functions
- Deep Reasoning: Suitable for tasks such as mathematical problems, complex logic problems, code reconstruction, and solution comparison and selection that require intermediate deduction chains.
- Code and Engineering Q&A: The model card mentions the enhanced experience of vibe coding, indicating that it is not only able to make up code, but is also better at continuous explanation and correction.
- Function Call: Compared with the previous version, function call support is more complete and suitable for hanging into workflow or tool agents for multi-step execution.
- File Understanding and Online Search: The web page has systematized file upload and search prompt word templates, indicating that the official is promoting the combination of "inference + external evidence".
- Local operation and distillation derivation: The model card also provides local operation instructions and discloses the results of distillation to Qwen3-8B, making it easier for the research team to migrate small models.
Expert view: Its most valuable thing is not a single function, but the linkage between "deep thinking + tool call + file/search context". In the past, many models could write code, but they would not reliably string external data, user uploaded content, and tool execution results into one package; R1 0528 is obviously making progress in this direction.
Model and version evolution
The product line of DeepSeek is already very clear: the V series is responsible for the general main force, and the R series is responsible for inference enhancement. The role of R1 0528 is to directly push the reasoning line R1 to a level that can replace some high-priced closed-source reasoning calls without renaming a new major version.
| Version | Date | Highlights of changes |
|---|---|---|
| DeepSeek-R1 | 2025-01-20 | Establishing the enhanced reasoning mainline for the first time |
| DeepSeek-V3-0324 | 2025-03-24 | Enhance general and coding capabilities to supplement the daily main line |
| DeepSeek-R1-0528 | 2025-05-28 | Inference depth, function calling, illusion control and coding experience are simultaneously enhanced |
Version logic: R1 0528 is not a product version that changes the UI, but an engineering upgrade that "makes the inference route more like a production model". For users, the most direct change is not the name, but fewer rollbacks and fewer rounds of prompt words in difficult scenarios.
Technical advantages
According to the main delivery form, DeepSeek R1 0528 belongs to [Basic Large Model/API Infrastructure]. Its advantage lies not in multi-modal work, but in turning high-quality reasoning into a low-cost, API-accessible, and open-source deployment capability base.
Performance and Throughput: The official has not publicly disclosed complete throughput data such as TTFT, TPM, RPM and so on for R1 0528. The current public API page mainly sells the V4 series, and R1 0528 is more suitable to be regarded as a benchmark of capabilities and an entrance to open source weight. The published concurrency information belongs to the V4 mainline. The specific online concurrency of R1 0528 is subject to the official real-time page.
Mechanism to Effect: Deeper thinking about token budget + post-training optimization directly leads to a jump in complex mathematics and programming benchmarks. The price is that some requests will be slower and the token will be more expensive, but the benefit is less rework and fewer second inquiries.
Adaptation Boundary: It is best at difficult mathematics, competition-level code, complex analysis, and tool calling links; it is least good at lightweight customer service that is extremely sensitive to delays, ultra-large-scale low-cost classification, and shallow Q&A that only requires a quick reply in one sentence.
API Example:
curl https://api.deepseek.com/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer <YOUR_API_KEY>" \
-d '{
"model": "deepseek-v4-flash",
"messages": [
{"role": "user", "content": "Please provide a complexity analysis of this Python code and point out potential bugs."}
],
"temperature": 0.2,
"max_tokens": 2048,
"stream": true,
"response_format": {"type": "text"}
}'The current official stable and verifiable V4 API name is used here; if the team wants to implement R1 0528 weight, the self-hosting and compatibility layer should be based on the model page and official real-time documents.
How to use
| Entrance | Suitable for the crowd | How to use |
|---|---|---|
| Web | Individual users, analysts, programmers | Log in to chat.deepseek.com, open DeepThink and ask questions directly |
| App | Heavy mobile users | Used for casual Q&A and document understanding, suitable for lightweight experience |
| API platform | Developers, product teams | Use OpenAI compatible interfaces to connect to existing systems |
| Local deployment | Research team, enterprise platform group | Download weights from Hugging Face and build your own inference service |
Recommended usage: First put it on "high-value, low-frequency" tasks, such as difficult bug analysis, complex financial logic explanation, contract comparison, and problem solving deduction. This is the easiest way to see the difference between it and the ordinary chat model.
It is not recommended to replace all the data as soon as it comes up: If the team transfers all customer service, search, and chat traffic to the inference model from the beginning, it will often only amplify the cost first, rather than gain value first.
Product Pricing
R1 0528 itself constitutes a "low-threshold entrance" with open source weight and web experience, and the official commercial price anchor is more reflected in the current document main sales model of the DeepSeek API platform.
| Project | Public Price |
|---|---|
| Web chat | Low threshold for personal experience |
| API input cache hit | Starting from $0.0028 / 1M tokens |
| API input cache miss | Starting from $0.14 / 1M tokens |
| API Output | Starting from $0.28 / 1M tokens |
| Concurrency limit | Mainline model ranges from 500 to 2500, depending on the model |
The truth about free: Free experience is suitable for testing abilities, but it does not mean free production. When actually doing batch reasoning, code scanning, and long text analysis, the main cost factors are the output token and concurrency quota.
Procurement Judgment: If the team has been able to stably use cheaper general models, R1 0528 is more suitable as an "upgrade file for difficult problems"; if a large number of the team's tasks are already on the edge of reasoning, then it may become the main model.
Application scenarios
- Dimensionality reduction attack scenario: Mathematical derivation, competition question explanation, complex code debugging, research-based Q&A, long-link problem troubleshooting, these tasks can best reflect its value.
- R&D collaboration scenario: The code reviewer first uses the ordinary model to do surface scanning, and then uses R1 0528 for troubleshooting, which can suppress both false positives and false negatives.
- Knowledge Work Scenario: When legal, investment research, and strategy positions handle highly complex texts, R1 0528 is more like a "second analyst" than a "writing assistant."
Dissuade scenario: High concurrency, extremely low customer service price; interaction that requires results within 1 second; rewriting and polishing that requires almost no reasoning; strong supervision and automatic decision-making that strictly prohibits any uncertain output.
Applicable people
- Individual researchers and programmers: Those who are willing to trade more waiting time for stronger reasoning quality will obviously feel the benefits.
- Developers and AI product teams: Suitable for use as a "difficult solution model" in difficult call chains.
- Enterprise Platform Group: If there is a model gateway, self-hosting and auditing system, the open source nature of R1 0528 will be more flexible than a purely closed source solution.
CURRENT LIMITATION: It is not the default answer for all processes. For scenarios with lightweight Q&A, extremely low latency, or a strong emphasis on unified format output, general models are often more cost-effective and easier to control.
Summary and Outlook
The core value of DeepSeek R1 0528 is to transform "difficult reasoning" from demonstration capabilities into model capabilities that are closer to real production. The most noteworthy aspect of it is not the single benchmark, but the advancement of deeper reasoning, function calling, file understanding and lower hallucination rates in the same version, which makes it very practical in R&D, research, and complex knowledge work.
There are two things you need to see clearly before purchasing or adopting: first, it saves the rework of difficult tasks and does not reduce costs in all scenarios; second, thinking deeply will eat up more tokens, which must be reserved for truly complex requests through routing strategies. A more reasonable way to implement it is to first put R1 0528 into the bottom layer of difficult tasks, observe the manual review rate, unit task cost and task completion rate, and then decide whether to expand the scope.
Related tools: 
DeepSeek, 
ChatGPT
Version Info
- DeepSeek-R1-0528 ：The official model card is defined as a minor version upgrade of DeepSeek R1. Through a deeper inference process and post-training algorithm optimization, it significantly enhances mathematics, programming, logic and function calling capabilities.
- DeepSeek-R1 ：R1 is the first formal reasoning mainline version, establishing the positioning of complex reasoning driven by reinforcement learning.
- DeepSeek-V3-0324 ：The enhanced version of the V3 main line focuses on general purpose and coding capabilities, forming another main product line besides R1.
User Reviews