DeepSeek-V3.2: benchmark labels distinguish AIME 2025 and Terminal-Bench 2.0

The OnAirToday AI Progress Tracker recorded a verified change to DeepSeek's DeepSeek-V3.2 (version V3.2): Benchmark labels distinguish AIME 2025 and Terminal-Bench 2.0.

DeepSeek-V3.2 is an open-weight language model supporting reasoning, coding and tool-enabled agent workflows. Its published evaluations show strong mathematical and software-engineering performance, while identifying remaining limitations in world knowledge and reasoning-token efficiency.

What changed in this review:

• Aider Polyglot and LMArena scores: Not yet confirmed by public reporting.
• January 2025 database exposure is recorded separately as a family-level infrastructure incident predating V3.2.
• Flaws and limitations updated
• Risks and threats updated

Teams building on DeepSeek-V3.2 should confirm limits, pricing and benchmark coverage against the vendor's own documentation before relying on these figures.

The full profile — capabilities, benchmark citations, known flaws and risk guidance — is on the tracker at /ai-models/deepseek-v3-2.

References:

• https://arxiv.org/abs/2512.02556
• https://huggingface.co/deepseek-ai/DeepSeek-V3.2
• https://api-docs.deepseek.com/quick_start/pricing
• https://www.wiz.io/blog/wiz-research-uncovers-exposed-deepseek-database-leak-sensitive-information