BigO(Bench) - Space Complexity Generation: leaderboard

Each model, zero-shot, must write a correct solution in each requested space-complexity class, checked by tests and the authors' profiling framework; all@1 counts a problem only when every class is met; each model ran on the 308 problems (636 problem-class pairs) of the test set, 20 samples each. These are the original protocol's numbers (paper v1/v2, March 2025, problem statements with their input bounds); the paper's 2026 v3 drops the bounds and re-measures, and is not this board. Llama 3.1 405B, which generated the input parsers that decide which problems enter the test set, is not ranked. Unlike HumanEval-style correctness boards it scores whether code meets a stated algorithmic complexity.

Metric: all@1 (%). Source: facebookresearch.github.io. Saturation forecast: Around 2029. 16 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1Llama 3.1 Nemotron Ultra 253B v15.6#1014
2QwQ-32B5.1#908
3DeepSeek R1 Distill Llama 70B3.4#1669
4DeepSeek R1 Distill Qwen 32B3.1#1806
5DeepSeek V3 (0324)2.5#722
6Llama 4 Scout Instruct2.2#1575
7Llama 3.3 70B Instruct1.8#1387
8Llama 3.1 70B Instruct1.8#1577
9DeepSeek V31.6#734
10Gemma 3 27B (IT)1.4#1339
11Codestral-22B-v0.11.3#2680
12Qwen 2.5 Coder 32B Instruct1.2#1597
13DeepSeek-Coder-V2-Instruct-07241#1103
14openhands-lm-32B-v0.10.9#2237
15Llama 4 Maverick Instruct FP80.8#1022

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=bigo-bench-space-complexity-generation · How It Works · Data refreshed daily, snapshot 2026-11-17.