DeepSeek V3.2: Open-Source AI Matching GPT-5 at One-Tenth the Cost
China's DeepSeek releases V3.2, a 671 billion parameter model that matches GPT-5 performance on reasoning benchmarks while costing 10x less. The entire model is MIT licensed. Here's what it means for the AI industry.
TL;DR - DeepSeek V3.2 Stats
- Parameters: 671B total, 37B activated per token (MoE)
- AIME 2025: 96.0% (GPT-5 High: 94.6%, Gemini 3 Pro: 95.0%)
- HMMT 2025: 99.2% accuracy
- Pricing: $0.028/M input tokens (~10x cheaper than GPT-5)
- License: MIT (fully open source)
- Training Cost: $6 million (vs GPT-4's $100 million)
The December 2025 Release
On December 1, 2025, China's DeepSeek unveiled two new versions of their experimental AI model: DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. According to Bloomberg, these models match the performance of OpenAI's flagship GPT-5 across multiple reasoning benchmarks.
December 2025 Model Lineup
Benchmark Performance
DeepSeek V3.2's performance on mathematical reasoning benchmarks is remarkable, matching or exceeding frontier models from OpenAI and Google:
AIME 2025 (Math Olympiad)
V3.2-Speciale Achievement: The advanced reasoning variant attains gold-level results in IMO (International Mathematical Olympiad), CMO (Chinese Mathematical Olympiad), ICPC World Finals, and IOI 2025.
The Cost Revolution
Perhaps more impressive than the benchmarks is DeepSeek's cost efficiency. According to Introl's analysis, DeepSeek offers frontier-level AI at a fraction of the cost:
Training Cost Comparison
Technical Architecture
DeepSeek V3 uses several innovative techniques to achieve its efficiency:
Mixture of Experts (MoE)
- 671B total parameters
- Only 37B activated per token
- Efficient sparse computation
Multi-head Latent Attention
- MLA architecture
- Reduced memory footprint
- Faster inference
Sparse Attention (V3.2)
- DeepSeek Sparse Attention
- More efficient attention mechanism
- Better long-context handling
Hybrid Thinking Mode
- Thinking and non-thinking modes
- Adaptive reasoning depth
- Cost-effective for simple queries
Getting Started with DeepSeek
import openai
# DeepSeek API is OpenAI-compatible
client = openai.OpenAI(
api_key="your-deepseek-api-key",
base_url="https://api.deepseek.com/v1"
)
response = client.chat.completions.create(
model="deepseek-chat", # or "deepseek-reasoner"
messages=[
{"role": "user", "content": "Solve this math problem..."}
]
)
print(response.choices[0].message.content)Self-Hosting: The full 671B parameter model is available on GitHub under the MIT license. You can run it locally with sufficient hardware.
Key Takeaways
- 1 Open source can compete: DeepSeek proves MIT-licensed models can match proprietary frontier AI.
- 2 Cost efficiency matters: 10x lower costs democratize access to frontier AI capabilities.
- 3 MoE is the future: Sparse activation enables huge models with practical inference costs.
- 4 China is competitive: Despite chip restrictions, Chinese AI labs are producing world-class models.
