Model Details
DeepSeek-V3-0324
Quick answer: DeepSeek-V3-0324 (released March 24, 2025) is an updated checkpoint of DeepSeek-V3 with targeted improvements to coding, reasoning, and instruction following. It raised SWE-bench Verified from 42.0% to 46.8% and GPQA from 59.1% to 64.8% over the original V3 release — maintaining the same MIT license, 685B MoE architecture, and $0.27/$1.10 API pricing.
What DeepSeek-V3-0324 Is
DeepSeek-V3-0324 is a dated checkpoint update to DeepSeek-V3, following DeepSeek's pattern of releasing improved model versions with date-stamped identifiers. The March 2024 update was notable for its coding improvements: the SWE-bench Verified score of 46.8% was the highest for any open-weight model at the time, and among the highest of any model at any price point.
The update also showed strong gains in science reasoning (GPQA Diamond: 64.8%) while maintaining V3's strengths in math and multilingual tasks. No architectural changes were made — this is a training improvement, not a new model.
Specifications
| Field | Value |
|---|---|
| Organization | DeepSeek |
| Total parameters | 685 billion (MoE) |
| Active parameters per token | 37 billion |
| Context window | 128,000 tokens |
| License | MIT |
| Release date | March 24, 2025 |
| HuggingFace | deepseek-ai/DeepSeek-V3-0324 |
Pricing
| Input (per 1M tokens) | Output (per 1M tokens) | |
|---|---|---|
| DeepSeek API | $0.27 | $1.10 |
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| SWE-bench Verified | 46.8% | DeepSeek blog | 2025-03 |
| GPQA Diamond | 64.8% | DeepSeek blog | 2025-03 |
| MATH-500 | 90.2% | DeepSeek blog | 2025-03 |
Scores reported by DeepSeek. Not Benchgen measurements.
Related Models & Benchmarks
- Compare: DeepSeek-V3 · DeepSeek-V3.1 · DeepSeek-R1-0528
- Benchmarks: SWE-bench Verified
Specs sourced from DeepSeek API docs. Last updated 2026-06-19.
DeepSeek-V3 0324 vs Alternatives
| Model | MMLU-Pro | BigCodeBench | AetherCode | License |
|---|---|---|---|---|
| DeepSeek-V3 0324 | 81.2% | 31.8% | 8.5% | MIT |
| DeepSeek-V3.1 | 83.7% | — | — | MIT |
| DeepSeek-V3.2 | — | — | — | MIT |
DeepSeek-V3 0324 (Mar 2025) → V3.1 (Jul 2025) improved MMLU-Pro from 81.2% to 83.7%. For coding: V3.1 LiveCodeBench (56.4%) likely better than 0324 AetherCode (8.5%).
Frequently Asked Questions
What is DeepSeek-V3 0324?
DeepSeek's March 24, 2025 model scoring 81.2% MMLU-Pro, 31.8% BigCodeBench. MIT — early V3 checkpoint.Scores from Benchgen evaluations. Last updated 2026-07-24.