To identify the performance of Large Language Models to solve problems with domain-specific knowledge (here: programming) we created a benchmark using the Project Euler series of challenging mathematical/computer programming problems. Because we want to apply LLMs for coding with specific programming languages, we also want to measure how good a LLM is at coding with that given programming language.
AI systems have domain-specific super-human perforances: chess-AI can compete against the best humans since 1997 where a chess system "Deep Blue" defeated the world champion Garry Kasparov. Since then the best chess programs are super-human and the same holds for most other games including Go, defeated by AlphaGo in 2016. So in domain-specific fields, AI programs have super-human performance and with the "Project Euler LLM Benchmark" we want to measure how strongly LLM models have super-human performances in the domain of coding or "being a programmer". See "Motivation" chapter below how we measure this.
The computed Benchmark ("PE-Bench-Python-100", "PE-Bench-Java-100", "PE-Bench-Rust-100", "PE-Bench-Clojure-100") is the super-human performance factor to code in Python/Java/Rust/Clojure. The "Economic Score" is the average performance per bytes of model size (times 100). Results are:
| Model | Best Model for Size (GB) |
PE-100- Score |
Mem- Score |
Size *10^9 Params |
Bits | Context Length (K) |
Python | Java | Rust | Clojure |
|---|---|---|---|---|---|---|---|---|---|---|
| DeepSeek-V3-0324 | 1342 | 17.06 | 1 | 671.0 | 16 | 128 | 20.08 | 18.63 | 14.91 | 4.62 |
| DeepSeek-V3 | 1342 | 16.11 | 1 | 671.0 | 16 | 128 | 20.01 | 16.95 | 12.16 | 5.92 |
| GPT-4o | 14.72 | 16 | 128 | 17.05 | 13.87 | 14.57 | 8.24 | |||
| GPT-o1-Mini | 200 | 11.05 | 6 | 100.0 | 16 | 32 | 17.44 | |||
| athene-v2:72b-q8_0 | 72.70 | 10.97 | 15 | 72.7 | 8 | 128 | 16.22 | 10.15 | 5.55 | 3.32 |
| hf.co/bartowski/Athene-V2-Agent-GGUF:Q4_K_M | 36.35 | 10.60 | 29 | 72.7 | 4 | 128 | 14.49 | 10.56 | 6.33 | 3.74 |
| athene-v2:latest | 36.35 | 10.44 | 29 | 72.7 | 4 | 128 | 14.07 | 11.14 | 6.55 | 1.62 |
| hf.co/bartowski/Sky-T1-32B-Preview-GGUF:Q4_K_M | 16.40 | 10.33 | 63 | 32.8 | 4 | 32 | 12.72 | 11.67 | 7.25 | 2.89 |
| hf.co/bartowski/Sky-T1-32B-Preview-GGUF:Q8_0 | 10.28 | 32.8 | 12.76 | 10.75 | 8.04 | 3.43 | ||||
| GPT-o1-Preview | 10.05 | 2 | 300.0 | 16 | 32 | 15.86 | ||||
| qwen2.5:72b-instruct-q4_K_M | 9.78 | 27 | 72.7 | 4 | 128 | 14.02 | 9.1 | 5.97 | 2.46 | |
| qwen2.5:72b-instruct-q8_0 | 9.77 | 13 | 72.7 | 8 | 128 | 12.98 | 10.5 | 5.41 | 3.49 | |
| qwen2.5-coder:32b-instruct-q4_K_M | 16.40 | 9.77 | 60 | 32.8 | 4 | 32 | 14.05 | 8.82 | 6.41 | 2.2 |
| hf.co/bartowski/Dracarys2-72B-Instruct-GGUF:Q4_K_M | 9.74 | 27 | 72.7 | 4 | 128 | 13.45 | 8.54 | 7.34 | 3.35 | |
| hf.co/mradermacher/Qwen2.5-72B-Instruct-abliterated-GGUF:Q4_K_M | 9.69 | 27 | 72.7 | 4 | 128 | 13.92 | 8.65 | 6.11 | 3.07 | |
| hf.co/bartowski/Sky-T1-32B-Flash-GGUF:Q4_K_M | 16.40 | 9.62 | 59 | 32.8 | 4 | 32 | 12.04 | 10.75 | 6.64 | 2.5 |
| hf.co/mradermacher/Bespoke-Stratos-32B-GGUF:Q4_K_M | 16.40 | 9.61 | 59 | 32.8 | 4 | 32 | 12.37 | 9.9 | 7.58 | 1.78 |
| qwen2.5-coder:32b-instruct-q8_0 | 9.38 | 29 | 32.8 | 8 | 32 | 11.94 | 9.94 | 6.69 | 2.8 | |
| qwen2.5:32b-instruct-q4_K_M | 16.40 | 9.17 | 56 | 32.8 | 4 | 32 | 12.34 | 9.26 | 6.13 | 2.25 |
| hf.co/bartowski/Qwen2.5-Coder-32B-Instruct-abliterated-GGUF:Q8_0 | 8.66 | 26 | 32.8 | 8 | 32 | 11.58 | 8.69 | 6.27 | 1.62 | |
| qwen2.5:32b-instruct-q8_0 | 8.44 | 26 | 32.8 | 8 | 32 | 9.73 | 10.22 | 5.91 | 2.98 | |
| GPT-4o-Mini | 8.00 | 16 | 128 | 11.39 | 7.36 | 5.19 | 1.93 | |||
| hf.co/bartowski/Qwen2.5-Coder-32B-Instruct-abliterated-GGUF:Q4_K_M | 16.40 | 7.95 | 48 | 32.8 | 4 | 32 | 9.55 | 9.3 | 5.71 | 1.94 |
| GPT-3.5-Turbo | 7.47 | 2 | 175.0 | 16 | 16 | 10.1 | 7.28 | 6.0 | 0.5 | |
| yi-coder:9b-chat-q8_0 | 8.80 | 7.37 | 84 | 8.8 | 8 | 128 | 11.08 | 6.77 | 4.32 | 0.47 |
| deepseek-coder:33b-instruct-q4_K_M | 7.34 | 44 | 33.0 | 4 | 16 | 9.55 | 10.72 | 0.0 | 3.03 | |
| llama3.3:70b-instruct-q8_0 | 7.17 | 10 | 70.6 | 8 | 128 | 8.93 | 8.06 | 4.29 | 3.17 | |
| qwen2.5-coder:14b-instruct-q8_0 | 7.09 | 48 | 14.8 | 8 | 128 | 9.7 | 7.35 | 4.55 | 0.95 | |
| hf.co/mradermacher/phi-4-abliterated-GGUF:Q8_0 | 6.81 | 14.7 | 10.06 | 6.67 | 3.68 | 0.52 | ||||
| hf.co/mradermacher/phi-4-abliterated-GGUF:Q4_K_M | 7.35 | 6.72 | 91 | 14.7 | 4 | 16 | 9.26 | 6.8 | 4.42 | 0.89 |
| qwen2.5-coder:14b-instruct-q4_K_M | 6.68 | 90 | 14.8 | 4 | 128 | 8.65 | 7.28 | 4.62 | 1.13 | |
| phi4:14b-q8_0 | 6.68 | 45 | 14.7 | 8 | 16 | 10.23 | 5.29 | 4.52 | 0.97 | |
| phi4:14b | 7.35 | 6.64 | 90 | 14.7 | 4 | 16 | 9.64 | 6.91 | 3.14 | 0.84 |
| phi4:latest | 7.35 | 6.64 | 90 | 14.7 | 4 | 16 | 9.64 | 6.91 | 3.14 | 0.84 |
| deepseek-coder:33b-instruct-q8_0 | 6.58 | 20 | 33.0 | 8 | 16 | 8.35 | 10.2 | 0.0 | 1.82 | |
| hf.co/mradermacher/Llama-3.1-SauerkrautLM-70b-Instruct-GGUF:Q4_K_M | 6.49 | 18 | 70.6 | 4 | 128 | 9.1 | 5.9 | 4.69 | 1.37 | |
| mistral-large:123b-instruct-2407-q4_K_M | 6.34 | 10 | 122.6 | 4 | 128 | 8.27 | 6.61 | 4.44 | 1.61 | |
| vanilj/Phi-4:Q8_0 | 6.13 | 42 | 14.7 | 8 | 16 | 9.06 | 5.73 | 3.52 | 0.84 | |
| yi-coder:9b-chat-q4_K_M | 4.40 | 5.97 | 136 | 8.8 | 4 | 128 | 7.44 | 6.04 | 5.76 | 0.34 |
| hf.co/bartowski/Anubis-70B-v1-GGUF:Q4_K_M | 5.83 | 17 | 70.6 | 4 | 128 | 8.07 | 6.49 | 2.59 | 1.36 | |
| llama3.1:70b-instruct-q8_0 | 5.81 | 8 | 70.6 | 8 | 128 | 8.19 | 5.36 | 3.8 | 1.7 | |
| qwen2.5:14b-instruct-q8_0 | 5.75 | 39 | 14.8 | 8 | 32 | 8.59 | 4.14 | 4.55 | 1.61 | |
| qwen2.5:14b-instruct-q4_K_M | 5.63 | 76 | 14.8 | 4 | 32 | 8.44 | 5.08 | 3.44 | 0.43 | |
| llama3.1:70b-instruct-q4_K_M | 5.62 | 16 | 70.6 | 4 | 128 | 8.77 | 4.98 | 2.77 | 0.6 | |
| llama3.3:70b-instruct-q4_K_M | 5.60 | 16 | 70.6 | 4 | 128 | 7.26 | 5.25 | 4.49 | 2.21 | |
| falcon3:10b-instruct-q8_0 | 5.48 | 53 | 10.3 | 8 | 32 | 8.15 | 5.14 | 2.71 | 1.39 | |
| hf.co/bartowski/smirki_UIGEN-T1.1-Qwen-14B-GGUF:Q4_K_M | 5.46 | 74 | 14.8 | 4 | 32 | 7.27 | 5.2 | 4.41 | 1.14 | |
| hf.co/mradermacher/calme-3.2-instruct-78b-GGUF:Q4_K_S | 5.43 | 14 | 78.0 | 4 | 32 | 7.04 | 6.05 | 3.6 | 0.74 | |
| tulu3:70b-q8_0 | 5.25 | 7 | 70.6 | 8 | 128 | 7.73 | 4.61 | 3.09 | 1.59 | |
| qwen2-math:72b-instruct-q8_0 | 5.25 | 7 | 72.7 | 8 | 4 | 6.49 | 6.67 | 2.61 | 1.3 | |
| tulu3:70b-q4_K_M | 5.21 | 15 | 70.6 | 4 | 128 | 7.31 | 4.68 | 3.35 | 2.15 | |
| hf.co/ozone-ai/0x-lite-Q4_K_M-GGUF:latest | 5.14 | 69 | 14.8 | 4 | 32 | 7.66 | 4.52 | 3.32 | 0.56 | |
| qwq:32b-preview-q8_0 | 5.09 | 16 | 32.8 | 8 | 32 | 9.68 | 2.94 | 1.39 | 0.54 | |
| command-a:111b-03-2025-q4_K_M | 5.06 | 9 | 111.1 | 4 | 256 | 6.33 | 5.17 | 4.21 | 1.32 | |
| hf.co/bartowski/Qwen2.5-14B-Instruct-1M-GGUF:Q4_K_M | 4.94 | 67 | 14.8 | 4 | 986 | 7.75 | 4.02 | 2.74 | 0.81 | |
| falcon3:10b-instruct-q4_K_M | 4.88 | 95 | 10.3 | 4 | 32 | 6.62 | 5.77 | 2.19 | 0.6 | |
| nemotron:70b-instruct-q4_K_M | 4.78 | 14 | 70.6 | 4 | 128 | 5.78 | 5.13 | 4.22 | 0.83 | |
| hf.co/bartowski/mistralai_Mistral-Small-3.1-24B-Instruct-2503-GGUF:Q4_K_M | 4.70 | 40 | 23.6 | 4 | 128 | 6.79 | 4.05 | 3.38 | 0.95 | |
| qwen2.5-coder:7b-instruct-q8_0 | 4.69 | 62 | 7.6 | 8 | 128 | 6.37 | 4.4 | 3.78 | 0.63 | |
| nemotron:70b-instruct-q8_0 | 4.50 | 6 | 70.6 | 8 | 128 | 6.31 | 4.05 | 3.23 | 1.13 | |
| qwen2.5:7b-instruct-q8_0 | 4.34 | 57 | 7.6 | 8 | 128 | 7.47 | 3.6 | 1.13 | 0.51 | |
| falcon3:7b-instruct-q8_0 | 4.34 | 58 | 7.5 | 8 | 32 | 6.76 | 3.91 | 2.16 | 0.36 | |
| hf.co/mradermacher/Viper-Coder-Hybrid-v1.3-GGUF:Q4_K_M | 4.32 | 58 | 14.8 | 4 | 128 | 6.44 | 4.45 | 1.83 | 0.46 | |
| hf.co/smirki/UIGEN-T1.1-Qwen-7B-Q4_K_M-GGUF:latest | 3.81 | 4.29 | 113 | 7.62 | 4 | 32 | 6.06 | 4.53 | 2.08 | 0.88 |
| qwen2.5:7b-instruct-q4_K_M | 3.80 | 4.22 | 111 | 7.6 | 4 | 128 | 7.5 | 2.67 | 1.86 | 0.49 |
| hf.co/mradermacher/Sky-T1-mini-GGUF:Q4_K_M | 4.01 | 105 | 7.62 | 4 | 128 | 7.05 | 3.68 | 0.15 | 0.55 | |
| mistral-small:24b-instruct-2501-q4_K_M | 3.98 | 34 | 23.6 | 4 | 32 | 6.32 | 2.89 | 2.47 | 0.93 | |
| qwen2.5-coder:7b-instruct-q4_K_M | 3.80 | 3.98 | 105 | 7.6 | 4 | 32 | 4.76 | 4.7 | 2.87 | 0.87 |
| hf.co/bartowski/open-thoughts_OpenThinker-32B-GGUF:Q4_K_M | 3.89 | 24 | 32.8 | 4 | 32 | 4.2 | 4.22 | 3.22 | 2.95 | |
| hf.co/bartowski/HelpingAI_Helpingai3-raw-GGUF:Q4_K_M | 3.70 | 72 | 10.3 | 4 | 32 | 5.88 | 3.82 | 0.92 | 0.13 | |
| gemma2:27b-instruct-q8_0 | 3.65 | 13 | 27.2 | 8 | 8 | 5.18 | 3.3 | 2.47 | 0.98 | |
| hf.co/bartowski/Athene-70B-GGUF:Q4_K_M | 3.58 | 10 | 70.6 | 4 | 8 | 6.98 | 1.99 | 0.76 | 0.36 | |
| hf.co/mradermacher/Viper-Coder-HybridMini-v1.3-GGUF:Q4_K_M | 3.55 | 93 | 7.62 | 4 | 32 | 5.18 | 3.51 | 1.99 | 0.24 | |
| deepseek-coder:6.7b-instruct-q8_0 | 3.52 | 50 | 7.0 | 8 | 16 | 5.37 | 3.68 | 0.94 | 0.79 | |
| hf.co/mradermacher/HelpingAI-3-GGUF:Q4_K_M | 3.38 | 66 | 10.3 | 4 | 32 | 5.8 | 2.82 | 1.09 | 0.0 | |
| gemma3:27b | 3.35 | 24 | 27.4 | 4 | 128 | 6.2 | 0.14 | 3.8 | 0.64 | |
| hf.co/bartowski/Yi-1.5-9B-Chat-GGUF:Q8_0 | 3.34 | 38 | 8.83 | 8 | 4 | 6.54 | 2.11 | 0.4 | 0.09 | |
| hf.co/lmstudio-community/Mistral-Small-24B-Instruct-2501-GGUF:Q4_K_M | 3.33 | 28 | 23.6 | 4 | 32 | 5.21 | 2.86 | 1.54 | 0.79 | |
| hf.co/internlm/internlm3-8b-instruct-gguf:Q4_K_M | 3.30 | 75 | 8.8 | 4 | 32 | 4.92 | 3.64 | 1.18 | 0.06 | |
| opencoder:8b-instruct-q8_0 | 3.21 | 41 | 7.8 | 8 | 8 | 4.63 | 3.22 | 1.62 | 0.72 | |
| hf.co/mradermacher/Bespoke-Stratos-7B-GGUF:Q4_K_M | 3.16 | 83 | 7.62 | 4 | 32 | 4.72 | 2.67 | 2.26 | 0.23 | |
| hf.co/bartowski/simplescaling_s1-32B-GGUF:Q4_K_M | 3.11 | 19 | 32.8 | 4 | 32 | 4.33 | 2.29 | 2.78 | 1.4 | |
| qwen2.5-coder:3b-instruct-q4_K_M | 1.55 | 3.05 | 197 | 3.1 | 4 | 32 | 4.98 | 2.51 | 1.53 | 0.03 |
| exaone3.5:32b-instruct-q8_0 | 3.05 | 10 | 32.0 | 8 | 32 | 3.96 | 3.82 | 1.38 | 0.47 | |
| deepseek-r1:32b-qwen-distill-q4_K_M | 3.00 | 18 | 32.8 | 4 | 128 | 4.51 | 2.91 | 1.44 | 0.32 | |
| hf.co/bartowski/Qwen2.5-7B-Instruct-1M-GGUF:Q4_K_M | 2.97 | 78 | 7.62 | 4 | 128 | 6.55 | 0.0 | 1.63 | 0.24 | |
| yi:34b-chat-v1.5-q4_K_M | 2.63 | 15 | 34.0 | 4 | 4 | 4.25 | 2.63 | 0.58 | 0.24 | |
| qwen2.5-coder:3b-instruct-q8_0 | 2.62 | 85 | 3.1 | 8 | 32 | 4.2 | 2.15 | 1.4 | 0.2 | |
| gemma3:12b | 2.55 | 42 | 12.2 | 4 | 128 | 4.62 | 0.41 | 2.59 | 0.58 | |
| yi:9b-chat-v1.5-q4_K_M | 2.54 | 56 | 9.0 | 4 | 4 | 4.04 | 2.71 | 0.39 | 0.33 | |
| hf.co/bartowski/Yi-1.5-34B-Chat-GGUF:Q8_0 | 2.45 | 7 | 34.4 | 8 | 4 | 4.61 | 1.49 | 0.72 | 0.16 | |
| exaone3.5:7.8b-instruct-q8_0 | 2.28 | 29 | 7.8 | 8 | 32 | 3.76 | 2.26 | 0.17 | 0.68 | |
| qwen:110b-chat-v1.5-q4_K_M | 2.25 | 4 | 111.0 | 4 | 32 | 3.63 | 1.84 | 0.76 | 0.92 | |
| phi3:14b-medium-128k-instruct-q8_0 | 2.24 | 16 | 14.0 | 8 | 128 | 4.21 | 1.55 | 0.42 | 0.04 | |
| llama3.1:8b-instruct-q8_0 | 2.03 | 25 | 8.0 | 8 | 128 | 3.26 | 1.78 | 0.94 | 0.09 | |
| tulu3:8b-q8_0 | 2.01 | 25 | 8.0 | 8 | 128 | 3.91 | 1.06 | 0.42 | 0.49 | |
| hf.co/bartowski/open-r1_OlympicCoder-7B-GGUF:Q4_K_M | 1.92 | 55 | 7.0 | 4 | 32 | 2.43 | 2.86 | 0.28 | 0.33 | |
| hf.co/bartowski/Yi-1.5-6B-Chat-GGUF:Q8_0 | 1.87 | 31 | 6.06 | 8 | 4 | 3.92 | 0.92 | 0.13 | 0.0 | |
| hf.co/bartowski/Yi-1.5-6B-Chat-GGUF:Q4_K_M | 1.81 | 60 | 6.06 | 4 | 4 | 3.71 | 0.87 | 0.32 | 0.0 | |
| deepseek-llm:67b-chat-q4_K_M | 1.79 | 5 | 67.0 | 4 | 4 | 2.94 | 1.63 | 0.5 | 0.23 | |
| qwq:32b-q4_K_M | 1.74 | 11 | 32.8 | 4 | 128 | 1.82 | 1.75 | 1.32 | 2.23 | |
| qwen2.5:3b-instruct-q4_K_M | 1.55 | 1.70 | 110 | 3.1 | 4 | 128 | 2.95 | 1.35 | 0.56 | 0.05 |
| deepseek-r1:70b-llama-distill-q4_K_M | 1.69 | 5 | 70.6 | 4 | 128 | 2.17 | 2.49 | 0.19 | 0.41 | |
| deepseek-r1:14b-qwen-distill-q4_K_M | 1.69 | 23 | 14.8 | 4 | 128 | 2.6 | 1.02 | 1.19 | 1.02 | |
| qwen2.5:3b | 1.55 | 1.67 | 108 | 3.1 | 4 | 128 | 2.95 | 1.35 | 0.42 | 0.05 |
| yi-coder:1.5b-chat-q4_K_M | 0.75 | 1.65 | 220 | 1.5 | 4 | 128 | 3.49 | 0.61 | 0.34 | 0.0 |
| gemma2:9b-instruct-q8_0 | 1.63 | 18 | 9.2 | 8 | 8 | 2.46 | 1.55 | 0.86 | 0.12 | |
| granite3.1-dense:8b-instruct-q8_0 | 1.59 | 19 | 8.2 | 8 | 128 | 2.73 | 1.55 | 0.16 | 0.03 | |
| granite3.2:8b-instruct-q4_K_M | 1.50 | 37 | 8.2 | 4 | 128 | 2.53 | 1.43 | 0.3 | 0.0 | |
| codegemma:7b-instruct-v1.1-q4_K_M | 1.46 | 33 | 9.0 | 4 | 8 | 2.21 | 1.49 | 0.66 | 0.01 | |
| exaone3.5:2.4b-instruct-q8_0 | 1.45 | 54 | 2.7 | 8 | 32 | 2.73 | 0.94 | 0.28 | 0.15 | |
| qwen2.5:3b-instruct-q8_0 | 1.43 | 46 | 3.1 | 8 | 128 | 2.87 | 0.59 | 0.44 | 0.18 | |
| deepseek-r1:7b-qwen-distill-q4_K_M | 1.43 | 38 | 7.6 | 4 | 128 | 2.64 | 1.08 | 0.0 | 0.46 | |
| opencoder:1.5b-instruct-q8_0 | 1.42 | 75 | 1.9 | 8 | 4 | 2.2 | 1.47 | 0.5 | 0.0 | |
| yi-coder:1.5b-chat-q8_0 | 1.36 | 91 | 1.5 | 8 | 128 | 2.32 | 1.17 | 0.42 | 0.0 | |
| qwen2.5-coder:1.5b-instruct-q4_K_M | 0.75 | 1.31 | 174 | 1.5 | 4 | 32 | 2.26 | 0.8 | 0.8 | 0.03 |
| mixtral:8x7b-instruct-v0.1-q4_K_M | 1.30 | 6 | 46.7 | 4 | 32 | 2.0 | 1.24 | 0.62 | 0.0 | |
| qwen2-math:7b-instruct-q8_0 | 1.28 | 17 | 7.6 | 8 | 4 | 2.49 | 0.95 | 0.02 | 0.0 | |
| dolphin3:8b-llama3.1-q8_0 | 1.27 | 16 | 8.0 | 8 | 128 | 2.3 | 0.89 | 0.26 | 0.31 | |
| qwen2.5-coder:1.5b-instruct-q8_0 | 1.22 | 81 | 1.5 | 8 | 32 | 2.04 | 0.9 | 0.66 | 0.03 | |
| hf.co/mradermacher/Dolphin3.0-Llama3.1-8B-abliterated-GGUF:Q8_0 | 1.21 | 15 | 8.03 | 8 | 128 | 2.29 | 0.47 | 0.58 | 0.37 | |
| codegemma:7b-instruct-q8_0 | 1.19 | 13 | 9.0 | 8 | 8 | 1.84 | 1.27 | 0.39 | 0.0 | |
| Bio-Medical-Llama-3-8B-GGUF:Q8_0 | 1.16 | 15 | 8.0 | 8 | 8 | 1.51 | 1.65 | 0.3 | 0.03 | |
| hf.co/mradermacher/HelpingAI-15B-GGUF:Q4_K_M | 1.16 | 15 | 15.3 | 4 | 4 | 2.42 | 0.64 | 0.0 | 0.0 | |
| falcon3:3b-instruct-q8_0 | 1.16 | 36 | 3.2 | 8 | 32 | 1.89 | 1.09 | 0.36 | 0.04 | |
| qwen2.5:1.5b-instruct-q8_0 | 1.14 | 76 | 1.5 | 8 | 128 | 2.07 | 0.82 | 0.29 | 0.06 | |
| deepseek-r1:8b-llama-distill-q4_K_M | 1.12 | 28 | 8.0 | 4 | 128 | 1.46 | 1.25 | 0.45 | 0.67 | |
| openchat:7b-v3.5-q8_0 | 1.02 | 15 | 7.0 | 8 | 8 | 1.62 | 1.21 | 0.06 | 0.0 | |
| codellama:34b-instruct-q4_K_M | 0.93 | 5 | 34.0 | 4 | 16 | 1.29 | 0.99 | 0.56 | 0.04 | |
| exaone-deep:7.8b-q4_K_M | 0.90 | 23 | 7.8 | 4 | 1.83 | 0.4 | 0.24 | 0.0 | ||
| qwen:32b-chat-v1.5-q4_K_M | 0.90 | 5 | 33.0 | 4 | 32 | 1.54 | 0.81 | 0.13 | 0.12 | |
| yi:6b-chat-v1.5-q4_K_M | 0.89 | 30 | 6.0 | 4 | 4 | 0.79 | 1.48 | 0.62 | 0.01 | |
| qwen2.5:1.5b-instruct-q4_K_M | 0.75 | 0.88 | 118 | 1.5 | 4 | 128 | 1.94 | 0.26 | 0.15 | 0.0 |
| gemma3:4b | 0.88 | 41 | 4.3 | 4 | 128 | 1.56 | 0.55 | 0.43 | 0.0 | |
| mixtral:8x7b-instruct-v0.1-q8_0 | 0.82 | 2 | 46.7 | 8 | 32 | 1.44 | 0.65 | 0.23 | 0.0 | |
| hf.co/bartowski/open-r1_OlympicCoder-32B-GGUF:Q4_K_M | 0.79 | 32.8 | 1.54 | 0.32 | 0.18 | 0.45 | ||||
| hf.co/mradermacher/HelpingAI-9B-GGUF:Q4_K_M | 0.78 | 18 | 8.83 | 4 | 4 | 1.64 | 0.41 | 0.0 | 0.0 | |
| hf.co/bartowski/OpenThinker-7B-GGUF:Q4_K_M | 0.75 | 20 | 7.62 | 4 | 32 | 0.88 | 0.77 | 0.53 | 0.62 | |
| exaone-deep:32b-q4_K_M | 0.70 | 4 | 32.0 | 4 | 32 | 0.93 | 0.66 | 0.36 | 0.53 | |
| command-r7b:7b-12-2024-q4_K_M | 0.68 | 17 | 8.0 | 4 | 128 | 1.54 | 0.03 | 0.26 | 0.04 | |
| llama3.2:3b | 0.66 | 41 | 3.2 | 4 | 128 | 1.42 | 0.18 | 0.21 | 0.0 | |
| olmo2:13b-1124-instruct-q4_K_M | 0.66 | 10 | 13.7 | 4 | 4 | 1.59 | 0.06 | 0.01 | 0.03 | |
| internlm2:7b-chat-v2.5-q4_K_M | 0.66 | 17 | 7.7 | 4 | 1.1 | 0.41 | 0.47 | 0.0 | ||
| hf.co/bartowski/agentica-org_DeepScaleR-1.5B-Preview-GGUF:Q4_K_M | 0.64 | 72 | 1.78 | 4 | 128 | 1.28 | 0.4 | 0.03 | 0.05 | |
| qwen:72b-chat-v1.5-q4_K_M | 0.64 | 2 | 72.0 | 4 | 32 | 0.94 | 0.78 | 0.13 | 0.0 | |
| olmo2:7b-1124-instruct-q4_K_M | 0.60 | 16 | 7.3 | 4 | 4 | 1.42 | 0.08 | 0.02 | 0.0 | |
| qwen2.5:0.5b-instruct-q8_0 | 0.50 | 0.57 | 115 | 0.5 | 8 | 128 | 1.33 | 0.0 | 0.21 | 0.0 |
| internlm2:1.8b-chat-v2.5-q4_K_M | 0.51 | 54 | 1.9 | 4 | 1.24 | 0.06 | 0.0 | 0.0 | ||
| granite3.1-dense:2b-instruct-q8_0 | 0.50 | 20 | 2.5 | 8 | 128 | 1.07 | 0.11 | 0.2 | 0.0 | |
| llama3.2:latest | 0.49 | 31 | 3.21 | 4 | 128 | 0.99 | 0.18 | 0.21 | 0.0 | |
| exaone-deep:2.4b-q4_K_M | 0.48 | 36 | 2.7 | 4 | 0.88 | 0.32 | 0.04 | 0.28 | ||
| codellama:13b-instruct-q4_K_M | 0.48 | 7 | 13.0 | 4 | 16 | 0.52 | 0.84 | 0.05 | 0.08 | |
| hf.co/trillionlabs/Trillion-7B-preview-GGUF:Q4_K_M | 0.47 | 13 | 7.53 | 4 | 4 | 0.98 | 0.09 | 0.21 | 0.11 | |
| smallthinker:3b-preview-q4_K_M | 0.47 | 27 | 3.4 | 4 | 128 | 0.98 | 0.25 | 0.0 | 0.0 | |
| deepseek-r1:1.5b-qwen-distill-q4_K_M | 0.46 | 51 | 1.8 | 4 | 128 | 0.87 | 0.24 | 0.03 | 0.37 | |
| granite3.2:2b-instruct-q4_K_M | 0.42 | 34 | 2.5 | 4 | 128 | 0.85 | 0.16 | 0.15 | 0.05 | |
| mistral:7b-instruct-q4_K_M | 0.41 | 12 | 7.0 | 4 | 32 | 0.48 | 0.71 | 0.04 | 0.0 | |
| smallthinker:3b-preview-q8_0 | 0.41 | 12 | 3.4 | 8 | 128 | 0.81 | 0.19 | 0.03 | 0.19 | |
| phi:2.7b-chat-v2-q4_K_M | 0.39 | 26 | 3.0 | 4 | 2 | 0.91 | 0.07 | 0.0 | 0.0 | |
| vicuna:33b-q4_K_M | 0.38 | 3 | 30.0 | 4 | 2 | 0.9 | 0.06 | 0.0 | 0.0 | |
| granite3.1-moe:3b-instruct-q8_0 | 0.35 | 11 | 3.3 | 8 | 128 | 0.8 | 0.03 | 0.11 | 0.03 | |
| hf.co/bartowski/UwU-7B-Instruct-GGUF:Q8_0 | 0.27 | 4 | 7.62 | 8 | 128 | 0.27 | 0.54 | 0.0 | 0.0 | |
| qwen2-math:1.5b-instruct-q8_0 | 0.27 | 18 | 1.5 | 8 | 4 | 0.65 | 0.03 | 0.0 | 0.0 | |
| gemma2:2b-instruct-q8_0 | 0.23 | 9 | 2.6 | 8 | 8 | 0.39 | 0.22 | 0.03 | 0.0 | |
| codellama:7b-instruct-q4_K_M | 0.21 | 6 | 7.0 | 4 | 16 | 0.33 | 0.03 | 0.33 | 0.01 | |
| deepseek-llm:7b-chat-q8_0 | 0.20 | 3 | 7.0 | 8 | 4 | 0.46 | 0.06 | 0.0 | 0.0 | |
| hf.co/QuantFactory/HelpingAI-3B-coder-GGUF:Q4_K_M | 0.20 | 11 | 3.48 | 4 | 128 | 0.31 | 0.22 | 0.0 | 0.1 | |
| llama3.2:1b | 0.18 | 15 | 1.2 | 8 | 0.41 | 0.06 | 0.0 | 0.0 | ||
| phi3:3.8b | 0.17 | 9 | 3.8 | 4 | 0.39 | 0.04 | 0.03 | 0.0 | ||
| hf.co/bartowski/Qwen2-VL-72B-Instruct-GGUF:Q4_K_M | 0.16 | 0 | 72.7 | 4 | 32 | 0.31 | 0.0 | 0.08 | 0.16 | |
| gemma3:1b | 0.50 | 0.15 | 30 | 1.0 | 4 | 32 | 0.13 | 0.33 | 0.0 | 0.0 |
| qwen2.5-coder:0.5b-instruct-q4_K_M | 0.25 | 0.15 | 60 | 0.5 | 4 | 32 | 0.36 | 0.01 | 0.01 | 0.0 |
| qwen2.5:0.5b-instruct-q4_K_M | 0.25 | 0.15 | 58 | 0.5 | 4 | 128 | 0.36 | 0.0 | 0.01 | 0.0 |
| qwen2.5-coder:0.5b-instruct-q8_0 | 0.14 | 28 | 0.5 | 8 | 32 | 0.33 | 0.01 | 0.03 | 0.0 | |
| llama2:70b-chat-q4_K_M | 0.14 | 0 | 69.0 | 4 | 4 | 0.31 | 0.03 | 0.01 | 0.03 | |
| llama3.2:1b-instruct-q8_0 | 0.13 | 11 | 1.2 | 8 | 128 | 0.28 | 0.06 | 0.0 | 0.0 | |
| falcon:7b-instruct-q4_0 | 0.11 | 3 | 7.0 | 4 | 2 | 0.28 | 0.0 | 0.0 | 0.0 | |
| falcon:40b-instruct-q4_0 | 0.11 | 1 | 40.0 | 4 | 2 | 0.07 | 0.0 | 0.42 | 0.0 | |
| phi4-mini:3.8b-q4_K_M | 0.11 | 6 | 3.8 | 4 | 128 | 0.23 | 0.04 | 0.01 | 0.0 | |
| falcon3:1b-instruct-q8_0 | 0.10 | 6 | 1.7 | 8 | 8 | 0.25 | 0.0 | 0.0 | 0.0 | |
| granite3.1-moe:1b-instruct-q8_0 | 0.10 | 7 | 1.3 | 8 | 128 | 0.24 | 0.0 | 0.0 | 0.0 | |
| qwen:14b-chat-v1.5-q4_K_M | 0.08 | 1 | 14.0 | 4 | 32 | 0.15 | 0.0 | 0.08 | 0.0 | |
| qwen:1.8b-chat-v1.5-q4_K_M | 0.06 | 6 | 2.0 | 4 | 32 | 0.0 | 0.0 | 0.0 | 0.62 | |
| vicuna:13b-q4_K_M | 0.04 | 1 | 13.0 | 4 | 2 | 0.09 | 0.01 | 0.0 | 0.0 | |
| granite3.2-vision:2b-q4_K_M | 0.04 | 3 | 2.5 | 4 | 16 | 0.07 | 0.03 | 0.0 | 0.0 | |
| vicuna:13b-v1.5-16k-q4_K_M | 0.03 | 1 | 13.0 | 4 | 2 | 0.06 | 0.03 | 0.0 | 0.0 | |
| qwen2:0.5b-instruct-q8_0 | 0.03 | 0 | 494.03 | 8 | 32 | 0.07 | 0.0 | 0.0 | 0.0 | |
| starcoder2:3b | 0.03 | 2 | 3.0 | 4 | 16 | 0.07 | 0.0 | 0.0 | 0.0 | |
| llama2:13b-chat-q4_K_M | 0.03 | 0 | 13.0 | 4 | 4 | 0.06 | 0.01 | 0.0 | 0.0 | |
| vicuna:7b-v1.5-16k-q4_K_M | 0.01 | 0 | 7.0 | 4 | 4 | 0.03 | 0.0 | 0.0 | 0.0 | |
| qwen:7b-chat-v1.5-q4_K_M | 0.01 | 0 | 8.0 | 4 | 32 | 0.02 | 0.0 | 0.0 | 0.03 | |
| deepseek-coder:1.3b-instruct-q8_0 | 0.01 | 1 | 1.0 | 8 | 16 | 0.0 | 0.02 | 0.0 | 0.0 | |
| llama2:7b-chat-q4_K_M | 0.00 | 0 | 7.0 | 4 | 4 | 0.01 | 0.0 | 0.0 | 0.0 | |
| smollm:1.7b-instruct-v0.2-q8_0 | 0.00 | 0 | 1.7 | 8 | 2 | 0.0 | 0.0 | 0.0 | 0.0 | |
| smollm:135m-instruct-v0.2-q8_0 | 0.14 | 0.00 | 0 | 0.135 | 8 | 2 | 0.0 | 0.0 | 0.0 | 0.0 |
| smollm:360m-instruct-v0.2-q8_0 | 0.00 | 0 | 0.36 | 8 | 2 | 0.0 | 0.0 | 0.0 | 0.0 | |
| qwen:0.5b-chat-v1.5-q4_K_M | 0.00 | 0 | 620.0 | 4 | 32 | 0.0 | 0.0 | 0.0 | 0.0 | |
| qwen:4b-chat-v1.5-q4_K_M | 0.00 | 0 | 4.0 | 4 | 32 | 0.0 | 0.0 | 0.0 | 0.0 | |
| codegemma:2b-code-v1.1-q4_K_M | 0.00 | 0 | 3.0 | 4 | 0.0 | 0.0 | 0.0 | 0.0 | ||
| vicuna:7b-q4_K_M | 0.00 | 0 | 7.0 | 4 | 0.0 | 0.0 | 0.0 | 0.0 |
This shows that even very small models like the llama3.2 model has a two-fold super-human performance at solving those problems.
Solving specific tasks by coding programs requires a high degree of accuracy and efficiency. Challenging problems, such as those presented by Project Euler, test whether a person (or an AI system) can comprehend complex problems and translate that understanding into effective solutions.
Because the Project Euler has statistics about the number of solved problems we have the ability to compute the likelihod for a human to solve the problem. The given counts of course also reflect several other causes for not solving the problem (not enough interest, not enough time, not visible at the time the contestant has subscribed to the project) which cannot be easily integrated into our measurement method; however Project Euler started 2015 with more than 160 problems so it is feasable to select only the first 100 problems for the PE-Bench-Python-100 benchmark.
We create the benchmark with the following concept:
- We have a fixed number of
participants = 1325386and a number of participants who solved each problemsolved_by - Each Problem has a specific likelihood to be solved, which is
percentage_solved = solved_by / participants(must be multiplied with 100 to get the percentage number) - Each Problem gets a number of points assigned that a participants gets as score, which is
points = participants / solved_by
So this leads to the effect, that an average person can solve each problem with an performance indicator of 1 (percentage_solved * points).
Therefore the average number of points for solving 100 problems is also 1. This gives us a nice baseline for human performance.
The data is scraped from the Project Euler web page and from other sources:
python3 problems_scraper.py
This loads all problems from https://projecteuler.net/ and stores them intoo the problems folder.
python3 solutions_scraper.py
This loads all solutions from https://raw.githubusercontent.com/lucky-bai/projecteuler-solutions/refs/heads/master/Solutions.md
and combines them with the number of solutions that users have submitted from https://projecteuler.net/archives
The result is stored in solutions.json. There also the data is enrichted with percentage_solved and points.
If we ask a LLM to solve 100 problems, add the achieved points for each correctly solved problem and divide it by 100, we get the performance
number for that LLM which compares to human performance by that average number. A performance value of 2 would mean "two-fold super-human performance".
In our test, small models like "llama3.2" already have that super-human performance.
A performance number over 1 indicates a super-human performance. It turns out that almost all LLMs have super-human performance.
The score used for the benchmark is computed in three steps:
For a given model and a given programming language, we loop over all (or 1..100) problems within the problems directory and perform the following task:
- load the problem description, i.e. from
problems/0001.txt - insert the problem description into a prompt template, i.e. from
templates/template_python.md - sending the resulting prompt to the selected LLM model using an api call to the ollama
/v1/chat/completionsendpoint - storing the answer from the model into `solutions/<model_name>//0001.md
This results in 100 answer files. This can be done calling
python3 inference.py --model <model_name>
Code is embedded into code blocks of the answer of the llm. We want to extract this in such a way, that a code interpreter can execute the file directly. This is done with the script
python3 codeextraction.py --model <model_name>
Finally the code is executed within a protected environment. This is done with
python3 execute.py --model <model_name>
where the python code is executed within. The resulting output line is truncated, only the last line is used. That line is compared
to the actual solution from the solutions.json file. The process does the code execution for all 100 problem solutions and adds up all the
points for the corresponding problem. This sum is divided by 100 and is the final score for the model.
As a preparation step for the tests, we must download the test cases from project euler with this script:
python3 problems_scraper.py
All over 900 tests are then stored in the problems folder. For our 100-bench-series, we use only the first 100.
To compute the benchmark, we use ollama as local inference engine.
To run a test (i.e. for the model athene-v2:72b-q8_0), the following command sequence is required:
python3 inference.py --language python --model athene-v2:72b-q8_0
python3 codeextraction.py --language python --model athene-v2:72b-q8_0
python3 execute.py --language python --model athene-v2:72b-q8_0
However, it is not convenient to do so, instead you can also call
python3 test.py --language python --model athene-v2:72b-q8_0
That computes all steps and updates the solutions.json file.
You can also call
python3 test.py --language python --allmodels --skip_existing
which loads the list of models from ollama and iterates over all models stored with ollama. This will take some time, if you have a large model collection in ollama maybe this takes longer than a week.
My current set-up does not allow to run models greater than 128GB. Models greater than this size had been tested using the API of the providing institution (i.e. OpenAI, DeepSeek). Other bigger models cannot be tested on my hardware.
Please send me a pull request for the following cases:
- If you have better hardware and want to contribute your contribution is welcome.
- The code runner for python, java, rust, clojure (see
execute.py) can possibly be enhanced. That would cause better benchmark scorings for affected models. Please see if this can be enhanced. - bugfixes (code and documentation)
This work (code and benchmark results) is licensed by Michael Christen under the Apache License Version 2.0, January 2004