DEV.co
Open-Source LLM · deepseek-ai

DeepSeek-R1-0528-Qwen3-8B

DeepSeek-R1-0528-Qwen3-8B is an 8.2B parameter open-source LLM distilled from DeepSeek-R1's reasoning capabilities into Qwen3's base model. It achieves strong performance on mathematics and coding benchmarks (AIME 2024: 86.0%, LiveCodeBench: 60.5%) while maintaining a compact footprint suitable for local deployment. MIT-licensed, ungated, and available on HuggingFace.

Source: HuggingFace — huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B
8.2B
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
1.6M
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerdeepseek-ai
Parameters8.2B
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads1.6M
Likes1.1k
Last updated2025-05-29
Sourcedeepseek-ai/DeepSeek-R1-0528-Qwen3-8B

What DeepSeek-R1-0528-Qwen3-8B is

Distilled reasoning model (Qwen3-8B base + DeepSeek-R1-0528 chain-of-thought post-training). Uses DeepSeek-R1-0528 tokenizer. Supports system prompts and reasoning patterns via <think> tokens. Max generation: 64K tokens. Evaluated at temperature 0.6, top-p 0.95. 1.56M downloads, last updated 2025-05-29.

Quickstart

Run DeepSeek-R1-0528-Qwen3-8B locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="deepseek-ai/DeepSeek-R1-0528-Qwen3-8B")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Mathematics and Competitive Problem-Solving

AIME 2024 performance (86.0%) and HMMT Feb 25 (61.5%) demonstrate strong reasoning on complex math. Suitable for educational tools, tutoring systems, and math competition preparation.

On-Device or Cost-Constrained Code Assistance

8B parameter size enables self-hosted deployment. LiveCodeBench 60.5% (2408-2505) sufficient for code review, refactoring suggestions, and junior-level coding tasks without cloud inference costs.

Research on Reasoning Distillation and Small Model Scaling

Explicitly designed as a research artifact demonstrating CoT distillation effectiveness. Valuable for academic work on model compression and reasoning capability transfer.

Running & fine-tuning it

ESTIMATE: 16–24GB VRAM for fp16 full inference; 8–12GB with 8-bit quantization; 4–6GB with 4-bit quantization (e.g., via llama.cpp, GPTQ). Exact requirements depend on sequence length and batch size. Verify with model card's local-run guidance or benchmarks before production deployment.

LoRA/QLoRA feasibility is plausible given 8B size and standard Qwen3 architecture, but model card does not explicitly document fine-tuning instructions. Uses DeepSeek-R1-0528 tokenizer (not original Qwen3), requiring careful config management. No official LoRA weights or adaptation guides provided in excerpt. Requires review of full repo.

When to avoid it — and what to weigh

  • Real-Time Latency-Critical Applications — Reasoning models generate 20K+ tokens per complex query (average 23K per AIME question). Inference speed unknown but likely slower than non-reasoning baselines. Not suitable for sub-second response SLAs.
  • Multi-Modal or Long-Context Requirements — Text-generation only; no vision capability. Context length unknown; no evidence of extended context window optimization beyond 64K max generation length.
  • Hallucination-Sensitive Production Systems — Model card notes 'reduced hallucination rate' vs. prior version but provides no metrics. SimpleQA score (27.8%) indicates factual accuracy gaps. Unsuitable for high-stakes factual retrieval without verification mechanisms.
  • Tasks Requiring Recent Knowledge or Web Access — Model evaluation frozen at 2025-05-29. No built-in web search or knowledge update mechanism documented. Card references web search prompts but does not claim integrated retrieval capability.

License & commercial use

MIT license. OSI-approved, permissive license allowing commercial use, modification, and redistribution with attribution. No copyleft restrictions.

MIT license explicitly permits commercial use, including closed-source applications and commercial services. Requires preservation of license notice and copyright attribution. No gating mechanism enforced. Commercial deployment is legally clear under MIT terms; however, practical deployment considerations (cost of inference, performance on specific tasks) remain independent of licensing.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityModerate
DEV.co fitGood
Assessment confidenceHigh
Security considerations

Card does not disclose security audits, red-teaming results, or adversarial robustness testing. Reasoning models with 20K+ token generations may amplify prompt injection or jailbreak effectiveness. No data provenance, synthetic data ratio, or filtering methodology documented. Use in sensitive contexts (e.g., medical, financial advice) requires independent safety evaluation. Self-hosted deployment removes third-party monitoring; apply standard LLM safety practices (input validation, output filtering, usage monitoring).

Alternatives to consider

Qwen3-8B (base, non-reasoning)

Lower inference latency; comparable general-knowledge performance (AIME 2024: 76% vs. 86%). Choose if reasoning overhead is unaffordable.

Phi-4-Reasoning-Plus-14B

Slightly smaller reasoning model (14B); AIME 2024: 81.3%, LiveCodeBench capable. Larger parameter count may offset inference cost savings.

Gemini 2.5 Flash Thinking (closed-source API)

Comparable or better benchmark performance; managed inference. Trade-off: no self-hosting, commercial API costs, vendor lock-in, but potentially lower latency and operational burden.

Software development agency

Ship DeepSeek-R1-0528-Qwen3-8B with senior software developers

DeepSeek-R1-0528-Qwen3-8B is MIT-licensed and immediately available on HuggingFace. Estimate your hardware needs, test quantization strategies, and check the GitHub repo for local-run examples. Devco can help integrate custom-tuning and optimized serving for your use case.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

DeepSeek-R1-0528-Qwen3-8B FAQ

Can I use this model commercially in a closed-source product?
Yes. MIT license permits commercial use, closed-source derivative works, and redistribution. You must retain copyright notices and the license text. No royalties or license fees apply.
What VRAM do I need to run this locally?
Estimated 16–24GB for fp16, 8–12GB for 8-bit quantization, or 4–6GB for 4-bit. The model card does not state exact requirements; verify with quantization tools (llama.cpp, bitsandbytes) and your target hardware before deployment.
Why does the model generate so many tokens for math problems?
The model is trained to perform explicit reasoning (chain-of-thought). On AIME, it averages 23K tokens per question to show its work, improving accuracy (87.5% vs. 70% in prior version). This is a feature, not a bug, but increases latency and cost per inference.
Can I fine-tune this model on my own data?
Likely yes, given 8B size and Qwen3 architecture, but the model card does not provide fine-tuning guides or pre-built LoRA weights. It uses a custom tokenizer (DeepSeek-R1-0528), so config management is critical. Requires experimental validation or review of the full GitHub repository.

Software development & web development with DEV.co

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If DeepSeek-R1-0528-Qwen3-8B is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.

Ready to Deploy a Reasoning-Capable 8B Model?

DeepSeek-R1-0528-Qwen3-8B is MIT-licensed and immediately available on HuggingFace. Estimate your hardware needs, test quantization strategies, and check the GitHub repo for local-run examples. Devco can help integrate custom-tuning and optimized serving for your use case.