MMODELYST
Evals/HumanEval
EVL

HumanEval

Python code synthesis

Source
benchmarkCodepass@1 (% solved correctly on the first try)164 items
What this measures

Can the model write correct Python? Given a function description, it must produce code that passes hidden unit tests. The standard first check of coding ability.

Sample
def has_close_elements(numbers, threshold): '''Return True if any two numbers are closer than threshold.''' — the model writes the body; it's run against tests.
Leaderboard · 16 models
pass@1 (% solved correctly on the first try) — higher is better
1Claude 3.5 Sonnet
92
2GPT-4o
90.2
3Grok-2
88.4
4Llama-3.3-70B-Instruct
88.4
5Claude 3.5 Haiku
88.1
6GPT-4o mini
87.2
7Qwen2.5-72B-Instruct
86.6
8Claude 3 Opus
84.9
9Qwen2.5-7B-Instruct
84.8
10DeepSeek-V3
82.6
11Llama-3.1-70B-Instruct
80.5
12Llama-3.1-8B-Instruct
72.6
13gemma-2-27b-it
71
14Phi-3.5-mini-instruct
62
15gemma-2-9b-it
60
16Mistral-7B-Instruct-v0.3
39
Benchmark health
active

Discriminating but climbing — the frontier is moving up this benchmark.

87.3
top-10 mean
92
best
16
models
What improves this score

Datasets & environments shown to raise it.