📡 LLM Pulse|🆕 New on leaderboard: Baidu: ERNIE-Light-DT-1.0 by Baidu — $1.13/M tokensLearn more →1/10

Measurement sources and verification status (12 fields)

Collected third-party figures are not independent tests. A source link and collection date do not establish reproducibility. Exact model version, dataset and test settings may be missing. No independent verification date is supplied.

speed: 80

latency: 0.6

contextWindow: 163840

inputPrice: 0.27

outputPrice: 0.41

benchmarkMmlu: 88.5

benchmarkHumanEval: 89.1

benchmarkMath: 86.8

benchmarkGpqa: 57

reasoningScore: 73

codingScore: 89.1

benchmarkSweBench: 49.2

DeepSeek: DeepSeek V3.2 Exp

deepseek🔓 Open Source

Published Humanity’s Last Exam leaderboard evidence.

LLM Stats Scores

73.0

Reasoning

89.1

Coding

🎯

57.0%GPQA Diamond — graduate-level science accuracy

Published benchmark evidence

Task scores are reported independently by each publisher; they are not combined into a hidden quality score.

View HLE source ↗

14.9%

Humanity's Last Exam

49.2%

SWE-bench

Quality Score

81.0

Overall (0–100)

Speed

80 c/s

characters/second

Latency (TTFT)

0.60s

to first token

Input Price

$0.27

per 1M tokens

Output Price

$0.41

per 1M tokens

Context Window

164K

tokens

Release Date

2025-09-29

30-Day Rank History

📊 Benchmark Scores

Reported scores · % accuracy

What do these benchmarks test?

🎯

GPQA Diamond57.0%Below avg

Designed so that even PhD-level human experts average ~65–70%. A score above 80% signals near-expert scientific reasoning that goes far beyond memorisation.

🧠

MMLU88.5%Top tier

The classic academic breadth benchmark. Top frontier models now score above 85%, so it best distinguishes mid-tier from strong generalist models.

💻

HumanEval89.1%Solid

Measures whether a model writes functionally correct code — not just plausible-looking syntax. Strong scores (>85%) indicate reliable day-to-day coding assistance.

∑

MATH Benchmark86.8%Top tier

Tests multi-step mathematical reasoning under strict correctness. Even a 60% score requires strong chain-of-thought; top reasoning models now exceed 90%.