DeepSeek-V3-0324 is a large-scale language model released by Depth Seeker, compared to its predecessor in thereasoning ability(e.g., MMLU-Pro improved from 75.9 to 81.2 and AIME jumped from 39.6 to 59.4),Front-end development code executability,Quality of Chinese writing(supports R1 style),search capability(Enhanced Reporting Analysis) andFunction Call Accuracyetc. were significantly optimized. The model parameters up to685BAdoptionMIT licenseThe template provides a temperature parameter mapping mechanism (API temperature 1.0 corresponds to model temperature 0.3) and a prompt template for file uploads/web searches.
Detailed summary
Model core information
| dimension (math.) | descriptive |
|---|---|
| name (of a thing) | DeepSeek-V3-0324 |
| publisher | DeepSeek (DeepSeek-AI) |
| parameter scale | 685B |
| authorization | MIT |
| Supported Features | Function Calls, JSON Output, FIM Completion, Multi-Language Support |
Key Improvements and Performance
- Improvement of reasoning skills
| benchmarking | Previous generation (DeepSeek-V3) | Current (V3-0324) | Enhancement |
|---|---|---|---|
| MMLU-Pro | 75.9 | 81.2 | +5.3 |
| GPQA | 59.1 | 68.4 | +9.3 |
| AIME | 39.6 | 59.4 | +19.8 |
| LiveCodeBench | 39.2 | 49.2 | +10.0 |
- functional enhancement
- front-end development: Optimize code executability and improve web / game interface aesthetics.
- Chinese Language Proficiency: Support for R1 writing styles, enhanced quality of mid-length content, optimized multi-round rewriting and translation capabilities.
- search capability: Enhance the detailed output of the report analysis.
- function call: Fix accuracy issues in previous versions.
Recommendations for use
- system alert: the date needs to be included, in the format
The assistant is DeepSeek Chat, created by Deep Seeker. Today is {current date}. - Temperature parameters: API temperature 1.0 corresponds to model temperature 0.3, it is recommended to call it through the mapping mechanism.
- Documentation / Search Tip Templates::
- The file upload template needs to contain the file name, content and question.
- Search results need to be combined with dates, filtered for relevance, and formatted to cite context.
Technical details
- model structure: Consistent with DeepSeek-V3, supports BF16, F8_E4M3, F32 precision.
- local deployment: Refer to the DeepSeek-V3 repository, Hugging Face Transformers is not supported at this time.
4. Key questions and answers
Q1: In what areas has DeepSeek-V3-0324 made significant improvements over its predecessor?
A1: Inreasoning ability(MMLU-Pro upgraded 5.3, AIME upgraded 19.8),Front-end development code executability,Quality of Chinese writing(supports R1 style),search capability(Enhanced Reporting Analysis) andFunction Call AccuracyAll aspects are significantly optimized.
Q2: What are the technical parameters and license of the model?
A2: The parameter scale is685BAdoptionMIT licenseThe DeepSeek-V3 repository supports BF16, F8_E4M3, and F32 accuracies, and requires local deployment through the DeepSeek-V3 repository.
Q3：How to call the model via API? What parameter settings do I need to pay attention to?
A3: When the API is called, the temperature parameter 1.0 will be mapped to the internal temperature of the model 0.3. You need to specify the date through the system prompts and follow the template format of the file upload and search prompts.
Priority Experience <strong>DeepSeek-V3-0324</strong> Click on the link below
- ￥Download for freeDownload after commentDownload after login
- {{attr.name}}:
