Direct answer
“DeepSeek 0528” usually refers to DeepSeek-R1-0528, the May 2025 update to DeepSeek-R1. The official model card reports stronger results than the earlier R1 on several reasoning, math, and coding benchmarks, but not every metric improved. It also reports a 64K maximum generation length for its evaluations; that is not, by itself, a specification of the model’s context window. DeepSeek-R1-0528 model card
For selection, distinguish the full model from DeepSeek-R1-0528-Qwen3-8B, which DeepSeek describes as a model distilled from R1-0528 into Qwen3 8B Base. Also distinguish both from a particular downloadable quantization, such as an Ollama listing. These are not interchangeable model identifiers. DeepSeek-R1-0528 model card
Production API verdict: Current DeepSeek documentation provides a live model-catalog endpoint, GET https://api.deepseek.com/models, intended to list currently available model IDs and metadata such as context and output limits. The documentation and quick-start reviewed here do not establish whether that catalog currently includes an exact deepseek-r1-0528 ID or a version pin. CometAPI’s page still contains a claim dated May 2025 that users could call deepseek-r1-0528, but the page does not document a current mapping of that ID to the full release or provide model-specific access conditions. So the exact hosted route remains unconfirmed here; check the provider’s current catalog and model documentation before integrating. DeepSeek model-catalog documentation DeepSeek API quick-start CometAPI page
What the name refers to
The official Hugging Face repository is named deepseek-ai/DeepSeek-R1-0528. Its model card presents it as an R1 update and publishes benchmark results and usage guidance. The repository’s displayed 689 GB size describes that repository listing, not a ready-to-run quantized download or a hardware requirement. DeepSeek-R1-0528 repository
The official card separately describes DeepSeek-R1-0528-Qwen3-8B as a distilled model: DeepSeek says it used R1-0528 chain-of-thought data to post-train Qwen3 8B Base. That lineage does not mean the 8B model has the same architecture, size, runtime needs, or performance as the full release. DeepSeek-R1-0528 model card
A community Ollama listing, for example, names sam860/deepseek-r1-0528-qwen3:8b and lists Qwen3 architecture, 8.19B parameters, Q4_K_M quantization, and a 5.0GB artifact. Those figures apply to that listed artifact—not automatically to the full model or every Qwen3-8B distribution. Ollama listing
Capability fit: what the published results support
The official model card compares R1-0528 with the earlier DeepSeek R1. Selected results include: DeepSeek-R1-0528 model card
| Evaluation | DeepSeek R1 | DeepSeek-R1-0528 | What it may indicate |
|---|---|---|---|
| AIME 2025 (Pass@1) | 70.0 | 87.5 | A substantial gain on this math competition benchmark |
| GPQA-Diamond (Pass@1) | 71.5 | 81.0 | Better reported performance on this science-question benchmark |
| LiveCodeBench, 2408–2505 (Pass@1) | 63.5 | 73.3 | Better reported performance on this coding benchmark range |
| SWE Verified (resolved) | 49.2 | 57.6 | Better reported task resolution under the stated evaluation setup |
| SimpleQA (correct) | 30.1 | 27.8 | A result that went down, despite broader claims of improvement |
The card says sampled benchmarks used temperature 0.6, top-p 0.95, and 16 responses per query to estimate Pass@1. Treat these results as evidence for those evaluations, not as a guarantee of quality on your prompts, languages, tools, or production data. DeepSeek-R1-0528 model card
A practical fit assessment: consider testing R1-0528 if your work depends on multi-step reasoning, math, or coding, since the official comparison reports gains on several benchmarks in those areas. Do not choose it solely on a general “better reasoning” claim: the published SimpleQA result is lower, and the card’s results do not replace evaluation on your own workload. Create a representative test set, score correctness and failure severity, and compare latency and cost using the exact model and serving configuration you expect to deploy.
The card reports a 64K maximum generation length for its evaluations. Do not treat that figure as the model’s context-window limit. DeepSeek-R1-0528 model card
Production API: hosted route versus self-hosting
A local or self-hosted endpoint is not the same thing as a hosted API product. The official model-card guidance includes vLLM and SGLang serving examples using the exact repository ID. For example, its vLLM instructions start a local server and call an OpenAI-compatible chat-completions endpoint: model usage and serving examples
pip install vllm
vllm serve "deepseek-ai/DeepSeek-R1-0528"
The card’s example then sends a request to http://localhost:8000/v1/chat/completions with deepseek-ai/DeepSeek-R1-0528 as the model. This documents a self-hosted setup pattern; it does not establish that the full weights will fit a particular machine, that the local service is production-hardened, or that DeepSeek hosts this model behind its API. model usage and serving examples
For hosted access, distinguish what the provider documents from what remains unknown:
- DeepSeek’s current API documentation: the quick-start describes the OpenAI-compatible API and names
deepseek-flashanddeepseek-v4-proin its sample configuration. The model-catalog documentation saysGET https://api.deepseek.com/modelslists currently available model IDs and metadata, including context-window and output limits. The documentation reviewed here does not itself establish whetherdeepseek-r1-0528is currently listed or provide a version pin. Query the catalog and verify the returned model ID and limits for your account before relying on an exact route. DeepSeek API quick-start DeepSeek model-catalog documentation - Historical first-party release note: DeepSeek’s May 2025 R1-0528 announcement said there was “no change to API usage” and linked the thinking-mode guide. That announcement documents the release-era API statement, not current availability of a pinned
deepseek-r1-0528identifier. DeepSeek R1-0528 release announcement - CometAPI: its page states that users could call
deepseek-r1-0528as of May 2025. Although the page is marked updated September 3, 2026, the captured current page still presents the R1-0528 access statement as a dated May 2025 claim; its general sample usesgpt-5-mini. The material reviewed here does not establish that CometAPI currently maps a model ID to the full R1-0528 release, nor does it establish model-specific endpoint behavior, limits, or access restrictions. Treat it as a lead to verify in CometAPI’s current model catalog and account—not as confirmation of a production route. CometAPI page
Before committing to any hosted provider, confirm in its current catalog, API documentation, and account settings:
- The exact model ID maps to full R1-0528, the Qwen3-8B derivative, or another variant—and whether that mapping is pinned or can move.
- The endpoint, request format, streaming behavior, and tool/function-calling support match your application.
- Authentication, availability for your account and intended region, rate limits, context and output limits, and service commitments suit your workload.
- If paid, the applicable currency and billing unit, and whether charges are per-token, subscription, or another arrangement.
No current R1-0528-specific hosted route or price is established by the provider material summarized here, so this guide does not recommend a provider. A provider’s model name alone is not enough to establish which weights or behavior it serves.
Selection takeaway
- Choose full R1-0528 as a candidate when your evaluation resembles reasoning, math, or coding tasks where its official card reports gains over R1—and validate it on your own data.
- Assess the Qwen3-8B derivative separately if you need a smaller artifact. It is a distinct distilled model, so test that exact distribution rather than borrowing full-model benchmark assumptions.
- For production, select the serving route as well as the model. Self-hosting examples exist. For hosted use, check the provider’s current model catalog and documentation for an exact R1-0528 mapping, version behavior, and operational limits; the dated CometAPI statement alone does not establish those facts.
The model card displays an MIT license label, but review the license terms and your use case before drawing legal conclusions. DeepSeek-R1-0528 model card
Sources
DeepSeek-R1-0528 model usage and serving examples
Ollama listing: sam860/deepseek-r1-0528-qwen3:8b
DeepSeek API list-models documentation