{ "server_url": "http://localhost:11434", "benchmarks": [ { "timestamp": "20250303_113430", "results": [ { "model": "wizard-vicuna-uncensored:latest", "total_duration": 12.927267513666667, "tokens_per_second": 23.964408789863477, "test_results": { "Fibonacci": { "success_rate": 0.0, "passed_cases": 0, "total_cases": 18, "success_cases_rate": 0.0, "avg_duration": 12.927267513666667, "avg_tokens_sec": 23.964408789863477 }, "Binary Search": { "success_rate": 0.0, "passed_cases": 0, "total_cases": 18, "success_cases_rate": 0.0, "avg_duration": 6.512057569666667, "avg_tokens_sec": 22.65644348402749 }, "Palindrome": { "success_rate": 0.0, "passed_cases": 0, "total_cases": 18, "success_cases_rate": 0.0, "avg_duration": 6.231747597333333, "avg_tokens_sec": 22.508357556533323 }, "Anagram Check": { "success_rate": 0.0, "passed_cases": 0, "total_cases": 18, "success_cases_rate": 0.0, "avg_duration": 10.538350027666667, "avg_tokens_sec": 23.50141235903869 } }, "overall_success_rate": 0.0, "overall_success_cases_rate": 0.0, "min_avg_duration": 6.231747597333333, "max_avg_duration": 12.927267513666667, "min_tokens_per_second": 22.508357556533323, "max_tokens_per_second": 23.964408789863477 }, { "model": "qwen2.5:14b", "total_duration": 40.78935244466666, "tokens_per_second": 8.889333441477179, "test_results": { "Fibonacci": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 40.78935244466666, "avg_tokens_sec": 8.889333441477179 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 48.256167, "avg_tokens_sec": 8.91781830131912 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 37.049296861, "avg_tokens_sec": 8.545959631111652 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 45.454861625, "avg_tokens_sec": 8.529429794367523 } }, "overall_success_rate": 100.0, "overall_success_cases_rate": 1.0, "min_avg_duration": 37.049296861, "max_avg_duration": 48.256167, "min_tokens_per_second": 8.529429794367523, "max_tokens_per_second": 8.91781830131912 } ] }, { "timestamp": "20250303_174821", "results": [ { "model": "wizard-vicuna-uncensored:latest", "total_duration": 9.061140611333332, "tokens_per_second": 22.0077609712149, "test_results": { "Fibonacci": { "success_rate": 0.0, "passed_cases": 0, "total_cases": 18, "success_cases_rate": 0.0, "avg_duration": 9.061140611333332, "avg_tokens_sec": 22.0077609712149 }, "Binary Search": { "success_rate": 0.0, "passed_cases": 0, "total_cases": 18, "success_cases_rate": 0.0, "avg_duration": 10.442756513666666, "avg_tokens_sec": 24.41797865756476 }, "Palindrome": { "success_rate": 33.33333333333333, "passed_cases": 6, "total_cases": 18, "success_cases_rate": 0.3333333333333333, "avg_duration": 5.603618597333333, "avg_tokens_sec": 22.36799915531131 }, "Anagram Check": { "success_rate": 5.555555555555555, "passed_cases": 1, "total_cases": 18, "success_cases_rate": 0.05555555555555555, "avg_duration": 11.451202444666666, "avg_tokens_sec": 23.838978284868205 } }, "overall_success_rate": 9.722222222222223, "overall_success_cases_rate": 0.09722222222222222, "min_avg_duration": 5.603618597333333, "max_avg_duration": 11.451202444666666, "min_tokens_per_second": 22.0077609712149, "max_tokens_per_second": 24.41797865756476 }, { "model": "qwen2.5-coder:7b-instruct-q4_K_M", "total_duration": 17.315668139, "tokens_per_second": 19.328937038685474, "test_results": { "Fibonacci": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 17.315668139, "avg_tokens_sec": 19.328937038685474 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 17.98539502766667, "avg_tokens_sec": 19.57693023406052 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 13.079861986, "avg_tokens_sec": 19.205164969270385 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 8.672462666666666, "avg_tokens_sec": 18.750156524185527 } }, "overall_success_rate": 100.0, "overall_success_cases_rate": 1.0, "min_avg_duration": 8.672462666666666, "max_avg_duration": 17.98539502766667, "min_tokens_per_second": 18.750156524185527, "max_tokens_per_second": 19.57693023406052 }, { "model": "samantha-mistral:latest", "total_duration": 12.211435388666667, "tokens_per_second": 23.923938967067834, "test_results": { "Fibonacci": { "success_rate": 44.44444444444444, "passed_cases": 8, "total_cases": 18, "success_cases_rate": 0.4444444444444444, "avg_duration": 12.211435388666667, "avg_tokens_sec": 23.923938967067834 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 7.592263986333333, "avg_tokens_sec": 24.005256761630953 }, "Palindrome": { "success_rate": 88.88888888888889, "passed_cases": 16, "total_cases": 18, "success_cases_rate": 0.8888888888888888, "avg_duration": 11.879134833, "avg_tokens_sec": 24.79139939867722 }, "Anagram Check": { "success_rate": 88.88888888888889, "passed_cases": 16, "total_cases": 18, "success_cases_rate": 0.8888888888888888, "avg_duration": 8.438326180666667, "avg_tokens_sec": 23.909059284433855 } }, "overall_success_rate": 80.55555555555556, "overall_success_cases_rate": 0.8055555555555556, "min_avg_duration": 7.592263986333333, "max_avg_duration": 12.211435388666667, "min_tokens_per_second": 23.909059284433855, "max_tokens_per_second": 24.79139939867722 }, { "model": "qwen2.5:14b", "total_duration": 35.248588736, "tokens_per_second": 9.780712468500422, "test_results": { "Fibonacci": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 35.248588736, "avg_tokens_sec": 9.780712468500422 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 30.093496236, "avg_tokens_sec": 9.852073250185546 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 31.317275361, "avg_tokens_sec": 9.878912443287856 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 34.83164913866667, "avg_tokens_sec": 9.864557197597287 } }, "overall_success_rate": 100.0, "overall_success_cases_rate": 1.0, "min_avg_duration": 30.093496236, "max_avg_duration": 35.248588736, "min_tokens_per_second": 9.780712468500422, "max_tokens_per_second": 9.878912443287856 }, { "model": "qwen2.5-coder:14b-instruct-q4_K_M", "total_duration": 37.17813055533333, "tokens_per_second": 9.679934868977442, "test_results": { "Fibonacci": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 37.17813055533333, "avg_tokens_sec": 9.679934868977442 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 23.064999083333333, "avg_tokens_sec": 9.768638447565982 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 29.162273486, "avg_tokens_sec": 9.648370262979713 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 34.809515569333335, "avg_tokens_sec": 9.880958976996757 } }, "overall_success_rate": 100.0, "overall_success_cases_rate": 1.0, "min_avg_duration": 23.064999083333333, "max_avg_duration": 37.17813055533333, "min_tokens_per_second": 9.648370262979713, "max_tokens_per_second": 9.880958976996757 }, { "model": "phi4:latest", "total_duration": 23.437334625000002, "tokens_per_second": 9.008754294006154, "test_results": { "Fibonacci": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 23.437334625000002, "avg_tokens_sec": 9.008754294006154 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 31.24377932, "avg_tokens_sec": 8.95641765835511 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 38.821998611333335, "avg_tokens_sec": 9.323287626758741 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 27.483618916666668, "avg_tokens_sec": 9.115156378168296 } }, "overall_success_rate": 100.0, "overall_success_cases_rate": 1.0, "min_avg_duration": 23.437334625000002, "max_avg_duration": 38.821998611333335, "min_tokens_per_second": 8.95641765835511, "max_tokens_per_second": 9.323287626758741 }, { "model": "openthinker:7b", "total_duration": 263.0048665136667, "tokens_per_second": 18.161543236382954, "test_results": { "Fibonacci": { "success_rate": 0.0, "passed_cases": 0, "total_cases": 18, "success_cases_rate": 0.0, "avg_duration": 263.0048665136667, "avg_tokens_sec": 18.161543236382954 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 168.908598889, "avg_tokens_sec": 18.286959395282295 }, "Palindrome": { "success_rate": 66.66666666666666, "passed_cases": 12, "total_cases": 18, "success_cases_rate": 0.6666666666666666, "avg_duration": 227.30224683333336, "avg_tokens_sec": 18.19731478522653 }, "Anagram Check": { "success_rate": 22.22222222222222, "passed_cases": 4, "total_cases": 18, "success_cases_rate": 0.2222222222222222, "avg_duration": 302.7885256943333, "avg_tokens_sec": 17.98162671601628 } }, "overall_success_rate": 47.22222222222222, "overall_success_cases_rate": 0.4722222222222222, "min_avg_duration": 168.908598889, "max_avg_duration": 302.7885256943333, "min_tokens_per_second": 17.98162671601628, "max_tokens_per_second": 18.286959395282295 }, { "model": "marco-o1:latest", "total_duration": 41.135754569666666, "tokens_per_second": 19.187625581350957, "test_results": { "Fibonacci": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 41.135754569666666, "avg_tokens_sec": 19.187625581350957 }, "Binary Search": { "success_rate": 33.33333333333333, "passed_cases": 6, "total_cases": 18, "success_cases_rate": 0.3333333333333333, "avg_duration": 51.497049624999995, "avg_tokens_sec": 19.299717479390168 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 33.27567922266667, "avg_tokens_sec": 19.39344271268353 }, "Anagram Check": { "success_rate": 88.88888888888889, "passed_cases": 16, "total_cases": 18, "success_cases_rate": 0.8888888888888888, "avg_duration": 47.74587681933333, "avg_tokens_sec": 19.37606411458147 } }, "overall_success_rate": 80.55555555555556, "overall_success_cases_rate": 0.8055555555555556, "min_avg_duration": 33.27567922266667, "max_avg_duration": 51.497049624999995, "min_tokens_per_second": 19.187625581350957, "max_tokens_per_second": 19.39344271268353 }, { "model": "mxbai-embed-large:latest", "total_duration": 0.0, "tokens_per_second": 0, "test_results": { "Fibonacci": { "success_rate": 0.0, "passed_cases": 0, "total_cases": 18, "success_cases_rate": 0.0, "avg_duration": 0.0, "avg_tokens_sec": 0 }, "Binary Search": { "success_rate": 0.0, "passed_cases": 0, "total_cases": 18, "success_cases_rate": 0.0, "avg_duration": 0.0, "avg_tokens_sec": 0 }, "Palindrome": { "success_rate": 0.0, "passed_cases": 0, "total_cases": 18, "success_cases_rate": 0.0, "avg_duration": 0.0, "avg_tokens_sec": 0 }, "Anagram Check": { "success_rate": 0.0, "passed_cases": 0, "total_cases": 18, "success_cases_rate": 0.0, "avg_duration": 0.0, "avg_tokens_sec": 0 } }, "overall_success_rate": 0.0, "overall_success_cases_rate": 0.0, "min_avg_duration": 0.0, "max_avg_duration": 0.0, "min_tokens_per_second": 0, "max_tokens_per_second": 0 }, { "model": "llama3.2:3b", "total_duration": 2.6735089863333332, "tokens_per_second": 36.09389841880591, "test_results": { "Fibonacci": { "success_rate": 77.77777777777779, "passed_cases": 14, "total_cases": 18, "success_cases_rate": 0.7777777777777778, "avg_duration": 2.6735089863333332, "avg_tokens_sec": 36.09389841880591 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 2.4100508473333333, "avg_tokens_sec": 35.29015421496925 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 1.0372564586666666, "avg_tokens_sec": 30.851865251357125 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 2.762424833333333, "avg_tokens_sec": 37.526840100741644 } }, "overall_success_rate": 94.44444444444444, "overall_success_cases_rate": 0.9444444444444444, "min_avg_duration": 1.0372564586666666, "max_avg_duration": 2.762424833333333, "min_tokens_per_second": 30.851865251357125, "max_tokens_per_second": 37.526840100741644 }, { "model": "llama3.2:1b-instruct-q4_K_M", "total_duration": 3.6423406526666664, "tokens_per_second": 88.23564288140822, "test_results": { "Fibonacci": { "success_rate": 27.77777777777778, "passed_cases": 5, "total_cases": 18, "success_cases_rate": 0.2777777777777778, "avg_duration": 3.6423406526666664, "avg_tokens_sec": 88.23564288140822 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 4.933635027666666, "avg_tokens_sec": 88.80398801189928 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 1.8736983196666668, "avg_tokens_sec": 88.92932647066883 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 1.938491514, "avg_tokens_sec": 88.64959146136897 } }, "overall_success_rate": 81.94444444444444, "overall_success_cases_rate": 0.8194444444444444, "min_avg_duration": 1.8736983196666668, "max_avg_duration": 4.933635027666666, "min_tokens_per_second": 88.23564288140822, "max_tokens_per_second": 88.92932647066883 }, { "model": "llama3.1:8b", "total_duration": 18.044723708666666, "tokens_per_second": 17.92435455357372, "test_results": { "Fibonacci": { "success_rate": 77.77777777777779, "passed_cases": 14, "total_cases": 18, "success_cases_rate": 0.7777777777777778, "avg_duration": 18.044723708666666, "avg_tokens_sec": 17.92435455357372 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 19.558708902666666, "avg_tokens_sec": 18.175417784298393 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 14.678415916666667, "avg_tokens_sec": 18.32735204911305 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 16.556909722, "avg_tokens_sec": 18.44864594957723 } }, "overall_success_rate": 94.44444444444444, "overall_success_cases_rate": 0.9444444444444444, "min_avg_duration": 14.678415916666667, "max_avg_duration": 19.558708902666666, "min_tokens_per_second": 17.92435455357372, "max_tokens_per_second": 18.44864594957723 }, { "model": "llama3.2-vision:11b-instruct-q4_K_M", "total_duration": 22.329897666666668, "tokens_per_second": 15.683487784271444, "test_results": { "Fibonacci": { "success_rate": 88.88888888888889, "passed_cases": 16, "total_cases": 18, "success_cases_rate": 0.8888888888888888, "avg_duration": 22.329897666666668, "avg_tokens_sec": 15.683487784271444 }, "Binary Search": { "success_rate": 94.44444444444444, "passed_cases": 17, "total_cases": 18, "success_cases_rate": 0.9444444444444444, "avg_duration": 28.854500333333334, "avg_tokens_sec": 14.92370109676282 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 16.310659278, "avg_tokens_sec": 15.077964284743096 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 19.173385736, "avg_tokens_sec": 15.919297403838554 } }, "overall_success_rate": 95.83333333333334, "overall_success_cases_rate": 0.9583333333333334, "min_avg_duration": 16.310659278, "max_avg_duration": 28.854500333333334, "min_tokens_per_second": 14.92370109676282, "max_tokens_per_second": 15.919297403838554 }, { "model": "hhao/qwen2.5-coder-tools:7b", "total_duration": 9.350419291333333, "tokens_per_second": 17.75326132604395, "test_results": { "Fibonacci": { "success_rate": 66.66666666666666, "passed_cases": 12, "total_cases": 18, "success_cases_rate": 0.6666666666666666, "avg_duration": 9.350419291333333, "avg_tokens_sec": 17.75326132604395 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 5.363153069333333, "avg_tokens_sec": 16.969020266567018 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 4.1743959723333335, "avg_tokens_sec": 16.053159652852415 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 4.8510482223333335, "avg_tokens_sec": 16.49191827040695 } }, "overall_success_rate": 91.66666666666666, "overall_success_cases_rate": 0.9166666666666666, "min_avg_duration": 4.1743959723333335, "max_avg_duration": 9.350419291333333, "min_tokens_per_second": 16.053159652852415, "max_tokens_per_second": 17.75326132604395 }, { "model": "falcon3:10b", "total_duration": 13.457690222666667, "tokens_per_second": 13.205079847530891, "test_results": { "Fibonacci": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 13.457690222666667, "avg_tokens_sec": 13.205079847530891 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 8.738829985999999, "avg_tokens_sec": 13.311836646406299 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 9.371235944333334, "avg_tokens_sec": 13.304170208123876 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 6.758179319666667, "avg_tokens_sec": 12.528827123222568 } }, "overall_success_rate": 100.0, "overall_success_cases_rate": 1.0, "min_avg_duration": 6.758179319666667, "max_avg_duration": 13.457690222666667, "min_tokens_per_second": 12.528827123222568, "max_tokens_per_second": 13.311836646406299 }, { "model": "deepseek-r1:8b", "total_duration": 168.97134605566666, "tokens_per_second": 17.431910724233774, "test_results": { "Fibonacci": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 168.97134605566666, "avg_tokens_sec": 17.431910724233774 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 107.91327369466667, "avg_tokens_sec": 18.012494102204858 }, "Palindrome": { "success_rate": 88.88888888888889, "passed_cases": 16, "total_cases": 18, "success_cases_rate": 0.8888888888888888, "avg_duration": 129.79005644466667, "avg_tokens_sec": 17.961536992747714 }, "Anagram Check": { "success_rate": 55.55555555555556, "passed_cases": 10, "total_cases": 18, "success_cases_rate": 0.5555555555555556, "avg_duration": 164.30645630566667, "avg_tokens_sec": 17.291549519011483 } }, "overall_success_rate": 86.11111111111111, "overall_success_cases_rate": 0.8611111111111112, "min_avg_duration": 107.91327369466667, "max_avg_duration": 168.97134605566666, "min_tokens_per_second": 17.291549519011483, "max_tokens_per_second": 18.012494102204858 }, { "model": "deepseek-r1:7b", "total_duration": 336.8736558336667, "tokens_per_second": 18.01216135231193, "test_results": { "Fibonacci": { "success_rate": 55.55555555555556, "passed_cases": 10, "total_cases": 18, "success_cases_rate": 0.5555555555555556, "avg_duration": 336.8736558336667, "avg_tokens_sec": 18.01216135231193 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 120.31249633333333, "avg_tokens_sec": 18.63905839923618 }, "Palindrome": { "success_rate": 66.66666666666666, "passed_cases": 12, "total_cases": 18, "success_cases_rate": 0.6666666666666666, "avg_duration": 331.0699081943333, "avg_tokens_sec": 18.283371808323412 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 78.71177402766666, "avg_tokens_sec": 19.069151610886887 } }, "overall_success_rate": 80.55555555555556, "overall_success_cases_rate": 0.8055555555555556, "min_avg_duration": 78.71177402766666, "max_avg_duration": 336.8736558336667, "min_tokens_per_second": 18.01216135231193, "max_tokens_per_second": 19.069151610886887 }, { "model": "deepseek-r1:1.5b-qwen-distill-q8_0", "total_duration": 137.589384375, "tokens_per_second": 57.36798365837254, "test_results": { "Fibonacci": { "success_rate": 61.111111111111114, "passed_cases": 11, "total_cases": 18, "success_cases_rate": 0.6111111111111112, "avg_duration": 137.589384375, "avg_tokens_sec": 57.36798365837254 }, "Binary Search": { "success_rate": 66.66666666666666, "passed_cases": 12, "total_cases": 18, "success_cases_rate": 0.6666666666666666, "avg_duration": 41.38244445833333, "avg_tokens_sec": 59.59776267442225 }, "Palindrome": { "success_rate": 33.33333333333333, "passed_cases": 6, "total_cases": 18, "success_cases_rate": 0.3333333333333333, "avg_duration": 371.12627313866665, "avg_tokens_sec": 53.88393896990123 }, "Anagram Check": { "success_rate": 50.0, "passed_cases": 9, "total_cases": 18, "success_cases_rate": 0.5, "avg_duration": 52.969490597, "avg_tokens_sec": 58.487717549099806 } }, "overall_success_rate": 52.77777777777778, "overall_success_cases_rate": 0.5277777777777778, "min_avg_duration": 41.38244445833333, "max_avg_duration": 371.12627313866665, "min_tokens_per_second": 53.88393896990123, "max_tokens_per_second": 59.59776267442225 }, { "model": "deepseek-r1:14b", "total_duration": 278.32433922266665, "tokens_per_second": 9.047140217393457, "test_results": { "Fibonacci": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 278.32433922266665, "avg_tokens_sec": 9.047140217393457 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 174.302479847, "avg_tokens_sec": 9.379116823142061 }, "Palindrome": { "success_rate": 88.88888888888889, "passed_cases": 16, "total_cases": 18, "success_cases_rate": 0.8888888888888888, "avg_duration": 482.098123792, "avg_tokens_sec": 9.218761695132402 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 308.77178886133333, "avg_tokens_sec": 8.899496740455705 } }, "overall_success_rate": 97.22222222222221, "overall_success_cases_rate": 0.9722222222222222, "min_avg_duration": 174.302479847, "max_avg_duration": 482.098123792, "min_tokens_per_second": 8.899496740455705, "max_tokens_per_second": 9.379116823142061 }, { "model": "Qwen2.5-Coder-7B-Instruct-s1k:latest", "total_duration": 9.948125833333332, "tokens_per_second": 18.377766002186945, "test_results": { "Fibonacci": { "success_rate": 88.88888888888889, "passed_cases": 16, "total_cases": 18, "success_cases_rate": 0.8888888888888888, "avg_duration": 9.948125833333332, "avg_tokens_sec": 18.377766002186945 }, "Binary Search": { "success_rate": 66.66666666666666, "passed_cases": 12, "total_cases": 18, "success_cases_rate": 0.6666666666666666, "avg_duration": 12.908918361333333, "avg_tokens_sec": 18.561003950076692 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 9.064120069666666, "avg_tokens_sec": 18.58546264084521 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 12.546619139, "avg_tokens_sec": 18.9448229322312 } }, "overall_success_rate": 88.88888888888889, "overall_success_cases_rate": 0.8888888888888888, "min_avg_duration": 9.064120069666666, "max_avg_duration": 12.908918361333333, "min_tokens_per_second": 18.377766002186945, "max_tokens_per_second": 18.9448229322312 }, { "model": "phi4-mini:latest", "total_duration": 10.860303611333332, "tokens_per_second": 29.361579428697542, "test_results": { "Fibonacci": { "success_rate": 61.111111111111114, "passed_cases": 11, "total_cases": 18, "success_cases_rate": 0.6111111111111112, "avg_duration": 10.860303611333332, "avg_tokens_sec": 29.361579428697542 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 10.22926025, "avg_tokens_sec": 29.360358027471495 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 7.7338954719999995, "avg_tokens_sec": 29.349959100715157 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 9.66612725, "avg_tokens_sec": 29.794841927435822 } }, "overall_success_rate": 90.27777777777779, "overall_success_cases_rate": 0.9027777777777778, "min_avg_duration": 7.7338954719999995, "max_avg_duration": 10.860303611333332, "min_tokens_per_second": 29.349959100715157, "max_tokens_per_second": 29.794841927435822 } ] }, { "timestamp": "20250313_051856", "results": [ { "model": "gemma3:12b", "total_duration": 17.904428624666668, "tokens_per_second": 11.206900603314153, "test_results": { "Fibonacci": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 17.904428624666668, "avg_tokens_sec": 11.206900603314153 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 14.096915041666666, "avg_tokens_sec": 11.209157987254114 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 9.514898375333333, "avg_tokens_sec": 11.037508677057549 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 24.419397555666666, "avg_tokens_sec": 11.87609409055045 } }, "overall_success_rate": 100.0, "overall_success_cases_rate": 1.0, "min_avg_duration": 9.514898375333333, "max_avg_duration": 24.419397555666666, "min_tokens_per_second": 11.037508677057549, "max_tokens_per_second": 11.87609409055045 } ] }, { "timestamp": "20250314_024439", "results": [ { "model": "SiliconBasedWorld/Qwen2.5-7B-Instruct-1M", "total_duration": 20.47047556933333, "tokens_per_second": 19.721316911932245, "test_results": { "Fibonacci": { "success_rate": 61.111111111111114, "passed_cases": 11, "total_cases": 18, "success_cases_rate": 0.6111111111111112, "avg_duration": 20.47047556933333, "avg_tokens_sec": 19.721316911932245 }, "Binary Search": { "success_rate": 66.66666666666666, "passed_cases": 12, "total_cases": 18, "success_cases_rate": 0.6666666666666666, "avg_duration": 89.59582123599999, "avg_tokens_sec": 19.522371869517652 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 29.476939527666666, "avg_tokens_sec": 19.835750358255293 }, "Anagram Check": { "success_rate": 33.33333333333333, "passed_cases": 6, "total_cases": 18, "success_cases_rate": 0.3333333333333333, "avg_duration": 52.099640236333336, "avg_tokens_sec": 19.661776969493513 } }, "overall_success_rate": 65.27777777777779, "overall_success_cases_rate": 0.6527777777777778, "min_avg_duration": 20.47047556933333, "max_avg_duration": 89.59582123599999, "min_tokens_per_second": 19.522371869517652, "max_tokens_per_second": 19.835750358255293 } ] }, { "timestamp": "20250314_110909", "results": [ { "model": "olmo2:13b", "total_duration": 25.239670416666666, "tokens_per_second": 8.973277631244137, "test_results": { "Fibonacci": { "success_rate": 61.111111111111114, "passed_cases": 11, "total_cases": 18, "success_cases_rate": 0.6111111111111112, "avg_duration": 25.239670416666666, "avg_tokens_sec": 8.973277631244137 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 10.511362861, "avg_tokens_sec": 8.094987124683419 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 7.803927528, "avg_tokens_sec": 8.07489922259982 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 16.829488430333335, "avg_tokens_sec": 8.85685146687769 } }, "overall_success_rate": 90.27777777777779, "overall_success_cases_rate": 0.9027777777777778, "min_avg_duration": 7.803927528, "max_avg_duration": 25.239670416666666, "min_tokens_per_second": 8.07489922259982, "max_tokens_per_second": 8.973277631244137 } ] }, { "timestamp": "20250314_111430", "results": [ { "model": "olmo2:13b-1124-instruct-q4_K_M", "total_duration": 27.796664694333334, "tokens_per_second": 9.16360668962085, "test_results": { "Fibonacci": { "success_rate": 27.77777777777778, "passed_cases": 5, "total_cases": 18, "success_cases_rate": 0.2777777777777778, "avg_duration": 27.796664694333334, "avg_tokens_sec": 9.16360668962085 }, "Binary Search": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 21.839994722333333, "avg_tokens_sec": 9.000336176480124 }, "Palindrome": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 10.587036805333334, "avg_tokens_sec": 8.492606444397637 }, "Anagram Check": { "success_rate": 100.0, "passed_cases": 18, "total_cases": 18, "success_cases_rate": 1.0, "avg_duration": 9.969617250333334, "avg_tokens_sec": 8.499243210997909 } }, "overall_success_rate": 81.94444444444444, "overall_success_cases_rate": 0.8194444444444444, "min_avg_duration": 9.969617250333334, "max_avg_duration": 27.796664694333334, "min_tokens_per_second": 8.492606444397637, "max_tokens_per_second": 9.16360668962085 } ] } ] }