DeepSeek-V3.2 Released: First Open-Source Model to Match GPT-5 with IMO/IOI Gold Medal Reasoning

DeepSeek has released DeepSeek-V3.2, the latest open-source large language model that achieves comprehensive breakthroughs in computational efficiency, reasoning capabilities, and agentic performance. DeepSeek-V3.2 matches GPT-5 in performance, while its high-compute variant DeepSeek-V3.2-Speciale surpasses GPT-5 and achieves reasoning proficiency on par with Gemini-3.0-Pro.

Key Technical Breakthroughs

1. DeepSeek Sparse Attention (DSA)

DeepSeek-V3.2 introduces the novel DeepSeek Sparse Attention (DSA) mechanism, an efficient attention architecture that substantially reduces computational complexity while preserving model performance.

DSA consists of two main components:

• Lightning Indexer: Computes index scores between query tokens and preceding tokens to determine which tokens should be selected
• Fine-grained Token Selection: Retrieves only the key-value entries corresponding to the top-k index scores

This design reduces core attention complexity from O(L²) to O(Lk), where k is much smaller than sequence length L. In 128K context scenarios, DSA achieves significant end-to-end speedup without noticeable performance degradation on long-context tasks.

2. Scalable Reinforcement Learning Framework

DeepSeek-V3.2 develops a stable and scalable RL protocol that enables significant computational expansion during post-training. Notably, the framework allocates a post-training computational budget exceeding 10% of pre-training cost, unlocking advanced capabilities.

The post-training pipeline includes:

• Specialist Distillation: Training specialized models for six domains including mathematics, programming, logical reasoning, and agentic tasks
• Mixed RL Training: Using GRPO algorithm to merge reasoning, agent, and human alignment training into a single RL stage
• Diverse Reward Mechanisms: Rule-based outcome rewards for reasoning and agent tasks, generative reward models for general tasks

3. Large-Scale Agentic Task Synthesis Pipeline

To integrate reasoning into tool-use scenarios, DeepSeek developed a novel synthesis pipeline that systematically generates training data at scale:

• Automatically synthesized 1,827 distinct environments and 85,000+ complex prompts
• Covers code agents, code interpreter agents, and general agents
• Synthetic tasks are “hard to solve but easy to verify,” ideal for reinforcement learning

Performance Benchmarks

Reasoning Capabilities

Benchmark	DeepSeek-V3.2-Thinking	DeepSeek-V3.2-Speciale	GPT-5-High	Gemini-3.0-Pro
AIME 2025 (Pass@1)	93.1%	96.0%	94.6%	95.0%
HMMT Feb 2025	92.5%	99.2%	88.3%	97.5%
LiveCodeBench	83.3%	88.7%	84.5%	90.7%
CodeForces Rating	2386	2701	2537	2708
HLE (Pass@1)	25.1%	30.6%	26.3%	37.7%

Competition-Level Performance

DeepSeek-V3.2-Speciale achieves gold-medal performance in top-tier mathematics and programming competitions:

Competition	Score	Medal
IMO 2025 (International Mathematical Olympiad)	35/42	🥇 Gold
CMO 2025 (China Mathematical Olympiad)	102/126	🥇 Gold
IOI 2025 (International Olympiad in Informatics)	492/600	🥇 Gold (10th place)
ICPC WF 2025 (ACM-ICPC World Finals)	10/12	🥇 Gold (2nd place)

Agentic Capabilities

DeepSeek-V3.2 significantly narrows the performance gap between open-source and closed-source models on agent tasks:

Benchmark	DeepSeek-V3.2	GPT-5-High	Claude-4.5-Sonnet
SWE-Verified (Resolved)	73.1%	77.2%	74.9%
Terminal Bench 2.0	35.2%	54.2%	42.8%
τ²-Bench (Pass@1)	80.3%	84.7%	80.2%
Tool Decathlon	35.2%	38.6%	29.0%

Inference Cost Advantages

Thanks to the DSA mechanism, DeepSeek-V3.2 significantly reduces inference costs in long-sequence scenarios:

• Prefilling Stage: At 128K token position, cost is approximately 1/3 of DeepSeek-V3.1-Terminus
• Decoding Stage: At 128K token position, cost is approximately 1/2 of DeepSeek-V3.1-Terminus

This makes DeepSeek-V3.2 a highly cost-effective choice for agentic scenarios.

Open Source & Deployment

DeepSeek-V3.2 continues DeepSeek’s open-source tradition, with model weights and inference code available on Hugging Face:

• Model Weights: https://huggingface.co/deepseek-ai/DeepSeek-V3.2
• Inference Code: https://huggingface.co/deepseek-ai/DeepSeek-V3.2/tree/main/inference

Technical Highlights Summary

1. Architecture Innovation: DSA sparse attention reduces long-sequence complexity from O(L²) to O(Lk)
2. Training Strategy: Post-training compute budget exceeds 10% of pre-training, using specialist distillation + mixed RL training
3. Data Synthesis: Automated pipeline generates 1,827 environments and 85,000+ complex tasks
4. Performance Breakthrough: First open-source model to achieve gold medals in IMO, IOI, and ICPC World Finals
5. Cost Optimization: 50-70% reduction in long-context inference costs

Future Directions

• Token Efficiency: DeepSeek-V3.2-Speciale’s token efficiency remains lower than Gemini-3.0-Pro
• Context Management: 128K context limit remains a bottleneck for some agentic tasks
• Self-Verification: Model tends to perform redundant self-verification, leading to excessively long trajectories

The release of DeepSeek-V3.2 marks the first time an open-source LLM has matched GPT-5 in comprehensive capabilities, providing the AI community with a powerful and efficient alternative. This demonstrates the viability of the open-source approach and opens new possibilities for future research and applications.

Related Links

• DeepSeek-V3.2 Technical Report
• Model Weights
• DeepSeek Official Website