Back to all models
D

DeepSeek-V2.5

DeepSeekReleased 2024-09-05
Official Page

DeepSeek's merged model combining deep reasoning with general instruction following at extremely low cost.

Overall Score
77%
Cost Efficiency
100
Context Window
128K

Strengths

  • ✓Extremely affordable
  • ✓Good coding performance
  • ✓Strong reasoning for price
  • ✓Open-source weights

Weaknesses

  • ✗Not as smart as frontier models
  • ✗MoE activation small
Architecture
Transformer (MoE)
Parameters
236B (21B active)
Input
$0.14/M
Output
$0.28/M

Benchmark Performance

BenchmarkDescriptionScoreCategoryBar
MMLUMassive Multitask Language Understanding86.8%knowledge
HumanEvalCode generation accuracy89.2%coding
MATHMathematical reasoning86%math
GSM8KGrade school math88%math
GPQAGraduate-level science Q&A49.7%reasoning
HellaSwagCommon sense reasoning85.3%reasoning
SWE-benchSoftware engineering problems36.4%coding
ARC-CAI2 Reasoning Challenge93%reasoning
AI
ModelBench

The open benchmark platform for comparing the latest AI models and their capabilities.

Platform

  • Compare Models
  • All Models
  • Methodology

AI Labs

  • OpenAI
  • Anthropic
  • Google DeepMind
  • Meta AI

Resources

  • Documentation
  • Benchmark Guide
  • API Reference
Benchmarks are sourced from official model cards and independent evaluations. Data is for informational purposes only.
© 2026 ModelBench. Built for the AI community.