DeepSeek · China · Released 2025-09

DeepSeek V3.2

Ultra-efficient MoE with sparse attention. Delivers frontier-adjacent quality at a fraction of the cost.

Official site GitHub

Access

open-weights

Context

128,000 tokens

Modalities

text

Languages

50+

Benchmark scores

MMLU

87.1%

Massive Multitask Language Understanding

GPQA

71.5%

Graduate-level science QA (Diamond)

HumanEval

92.7%

Python code completion

SWE-bench

49.2%

Real GitHub issue resolution (Verified)

LiveCodeBench

N/A

Contamination-free coding

AIME

N/A

American Invitational Mathematics Exam

MMMU

N/A

Multimodal college-level reasoning

MATH

90.2%

Competition-level math word problems

Scores as reported by the vendor or leading public leaderboards. "N/A" means the score has not been publicly disclosed for this metric.

Explain like I'm 5

Almost as smart as the top models but way cheaper to run.

Key features

Sparse attention
MoE
Very low pricing
Open weights

Strengths

Extreme cost efficiency
Strong coding
Open weights

Limitations

Text-only in flagship variant

Best for

High-volume APIsCoding copilotsSelf-hosting

Related models

OpenAI

GPT-5

OpenAI's flagship reasoning + multimodal model with a unified router that switches between fast and deep-thinking modes.

Anthropic

Claude Sonnet 4.5

Anthropic's best coding + agentic model, purpose-built for long-horizon computer-use and software engineering tasks.

Google DeepMind

Gemini 2.5 Pro

Google's most capable model with native 1M-token context and full multimodal I/O across text, images, audio and video.

xAI

Grok 4

xAI's frontier reasoning model with real-time X data grounding and a heavy multi-agent variant (Grok 4 Heavy).