MMODELYST
Evals/LiveCodeBench
EVL

LiveCodeBench

Contamination-free coding

Source ↗
benchmarkCodepass@1 on held-out, time-stamped problems700 items
What this measures

Competitive-programming problems collected *after* models were trained, so they can't have memorized the answers — a contamination-free measure of real coding skill.

Sample
A recent LeetCode/Codeforces-style problem (e.g., 'find the longest path in a weighted DAG') with full test suites.
Leaderboard · 98 models
pass@1 on held-out, time-stamped problems — higher is better
1Gemini 3 Pro Preview
91.7
2Gemini 3 Flash Preview
90.8
3DeepSeek V3.2 Speciale
89.6
4GLM-4.7
89.4
5GPT-5.2
88.9
6gpt-oss-120b
87.8
7Claude Opus 4.5
87.1
8GPT-5.1
86.8
9MiMo-V2-Flash
86.8
10DeepSeek V3.2
86.2
11o4-mini
85.9
12Kimi K2 Thinking
85.3
13GPT-5.1 Codex
84.9
14GPT-5
84.6
15GPT-5 Codex
84
16GPT-5.1 Codex mini
83.6
17Grok 4 Fast
83.2
18MiniMax-M2
82.6
19Grok 4.1 Fast
82.2
20Grok 4
81.9
21ERNIE 5.0 Thinking Preview
81.2
22MiniMax-M2.1
81
23o3
80.8
24Apriel-v1.6-15B-Thinker
80.7
25Gemini 2.5 Pro
80.1
26DeepSeek V3.1 Terminus
79.8
27DeepSeek V3.2 Exp
78.9
28GPT-5 nano
78.9
29Qwen3 235B A22B 2507
78.8
30Qwen3 Next 80B A3B
78.4
31Gemini 2.5 Pro Preview (Mar' 25)
77.8
32INTELLECT-3
77.7
33gpt-oss-20b
77.7
34DeepSeek R1 0528 (May '25)
77
35Gemini 2.5 Pro Preview (May' 25)
77
36K-EXAONE
76.8
37Qwen3 Max
76.7
38Doubao Seed Code
76.6
39Seed-OSS-36B-Instruct
76.5
40Magistral Medium 1.2
75
41KAT-Coder-Pro V1
74.7
42EXAONE 4.0 32B
74.7
43NVIDIA Nemotron 3 Nano 30B A3B
74.1
44Qwen3 VL 32B
73.8
45GLM-4.5
73.8
46Llama Nemotron Super 49B v1.5
73.7
47Nova 2.0 Pro Preview
73
48Apriel-v1.5-15B-Thinker
72.8
49NVIDIA-Nemotron-Nano-9B-v2
72.4
50Falcon-H1R-7B
72.4
51Magistral Small 1.2
72.3
52o3-mini
71.7
53Claude 4.5 Sonnet
71.4
54Gemini 2.5 Flash Preview (Sep '25)
71.3
55MiniMax M1 80k
71.1
56Qwen3 30B A3B 2507
70.7
57Qwen3 VL 30B A3B
69.7
58Grok 3 mini Reasoning
69.6
59GLM-4.6
69.5
60Gemini 2.5 Flash
69.5
61Olmo 3.1 32B Think
69.5
62K2-V2
69.4
63NVIDIA Nemotron Nano 12B v2 VL
69.4
64GPT-5 mini
69.2
65Gemini 2.5 Flash-Lite Preview (Sep '25)
68.8
66Cogito v2.1
68.8
67Hermes 4 - Llama-3.1 405B
68.6
68GLM-4.5-Air
68.4
69Qwen3 Next 80B A3B Instruct
68.4
70o1
67.9
71Qwen3 Omni 30B A3B
67.9
72Ling-1T
67.7
73Olmo 3 32B Think
67.2
74Nova 2.0 Lite
66.3
75Nova 2.0 Omni
66
76MiniMax M1 40k
65.7
77Grok Code Fast 1
65.7
78Mi:dm K 2.5 Pro
65.6
79Claude 4 Sonnet
65.5
80Claude 4.1 Opus
65.4
81Hermes 4 - Llama-3.1 70B
65.3
82Qwen3 Max (Preview)
65.1
83Motif-2-12.7B-Reasoning
65.1
84Qwen3 VL 235B A22B
64.6
85Ring-1T
64.3
86Qwen3 4B 2507
64.1
87Llama 3.1 Nemotron Ultra 253B v1
64.1
88Claude 4 Opus
63.6
89QwQ 32B
63.1
90HyperCLOVA X SEED Think (32B)
62.9
91Ring-flash-2.0
62.8
92Qwen3 235B A22B
62.2
93DeepSeek R1 (Jan '25)
61.7
94Olmo 3 7B Think
61.7
95Solar Pro 2
61.6
96Claude 4.5 Haiku
61.5
97Kimi K2 0905
61
98GLM-4.5V
60.4
Benchmark health
saturating

Top models cluster near the ceiling — differences here are getting too small to mean much. Weight newer, harder benchmarks more.

88.5
top-10 mean
91.7
best
98
models
Recent changes
GPT-5 mini83.8 → 69.210d ago
Nova 2.0 Lite71.1 → 66.317d ago
What improves this score

Datasets & environments shown to raise it.

SWE-Gym+7.0
The Stack v2+6.0
Research using this benchmark
all →
VibeThinker-3B: Exploring the Frontier of Verifiable Reasoning in Small Language ModelsMulti-LCB: Extending LiveCodeBench to Multiple Programming LanguagesBenchEvolver: Frontier Task Synthesis via Solution-Centric EvolutionREVES: REvision and VErification--Augmented Training for Test-Time ScalingQwen2 Technical Report

Matched by name in title/abstract.