DeepSeek-V3.2-Speciale is a high-compute variant of DeepSeek-V3.2 optimized for maximum reasoning and agentic performance. It builds on DeepSeek Sparse Attention (DSA) for efficient long-context processing, then scales post-training reinforcement learning...
Tested on 9 benchmarks · BenchGecko score 95.2. Top scores: OpenCompass — AIME2025 (96.0%), OpenCompass — IFEval (91.7%), OpenCompass — GPQA-Diamond (86.7%).
Step 3.5 Flash scores 89.5 (94% as good) at $0.10/1M input · 75% cheaper
OpenCompass Live Code Bench v6. Fresh competitive programming problems to evaluate code generation without memorization.
OpenCompass evaluation on AIME 2025 problems. Tests mathematical reasoning on fresh competition problems.
OpenCompass evaluation of GPQA Diamond. PhD-level science questions from the hardest subset.
OpenCompass MMLU-Pro evaluation. Harder knowledge test with more answer choices.
OpenCompass evaluation of Humanitys Last Exam. Expert-level cross-discipline knowledge test.
- Typetext
- Context164K tokens (~82 books)
- ReleasedDec 2025
- LicenseOpen Source
- StatusActive
- Cost / Message~$0.002
Frequently Asked Questions
Related Models
GPT-6 Astra · OpenAIClaude Opus 5.5 · AnthropicClaude Mythos Preview · AnthropicGPT-5.5 Pro · OpenAIStep 3.5 Flash · stepfunKey facts · as of 2026-05-03
- DeepSeek V3.2 Speciale by DeepSeek. BenchGecko score 95.2, rank 5 of 312 scored models (normalized average of public benchmark scores).
- List price $0.40 input · $1.20 output per 1M tokens (as of 2026-05-03).
How to cite · data as of 2026-05-03
DeepSeek V3.2 Speciale · benchmarks, pricing and providers. BenchGecko, data as of 2026-05-03. https://benchgecko.ai/model/deepseek-v3-2-speciale
Credit "Source: BenchGecko" with a link. Prices per provider and Gecko Tests are BenchGecko data (CC BY 4.0); benchmark scores keep their original source, listed in the JSON. JSON · llms.txt · MCP