DeepSeek V3.2 Exp (Reasoning) AI Evaluation | LayerLens

Model

DeepSeek V3.2 Exp (Reasoning) by DeepSeek

text->text

Free credits included. No credit card required.

DeepSeek-V3.2-Exp is an experimental large language model released by DeepSeek as an intermediate step between V3.1 and future architectures. It introduces DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism designed to improve training and inference efficiency in long-context scenarios while maintaining output quality.

Evaluations

across 14 benchmarks

Latency

4m 27s

Context length

164k

tokens

Cost

$0.28 · $0.41

input · output per 1M tokens

Key takeaways

DeepSeek V3.2 Exp (Reasoning) is an experimental Transformer-based LLM focused on improving efficiency in long-context scenarios using DeepSeek Sparse Attention (DSA).
The model excels in mathematical reasoning, achieving 96.2% accuracy on the MATH-500 dataset for multi-choice math problems and providing detailed explanations.
A significant limitation is its poor performance in complex medical and biological reasoning tasks, as evidenced by only 11.22% accuracy on 'Humanity's Last Exam'.
The model struggles with applying general knowledge, misinterpreting questions, and performing multi-step reasoning in domain-specific areas like medicine.
It frequently generates lengthy, overconfident, and incorrect explanations, especially in biological/medical contexts, indicating a lack of robust general knowledge application and overconfidence in wrong answers.

Technical specifications

Context length: 164k tokens
Max output: —
Architecture: Transformer
Modalities: text->text
Parameters: ~800B
Open weights: Yes
Licence: Open
Regions: China
Vendor: DeepSeek
Provider model key: deepseek/deepseek-v3.2-exp-reasoning