llmrun

llmrun Composite Score · Coding

The best LLMs for coding, open and proprietary together, on one open-ended scale where 100 is the average model (coding). Switch to open models to see only what you can download and run.

46 benchmarks4 sourcesUpdated 14 Nov 2026Reference scale 2026-Q4How the Composite Score is built

Coding score vs VRAM needed

This chart doesn't show proprietary models.

Models ranked by coding llmrun Composite Score
#Modelllmrun Composite ScoreCodingAgentsMathScienceReasoningBenchmarksVRAM
121MiMo v2 Flashopen187 GB · 3 benchmarks—Coding 100Agents –Math –Science 83Reasoning –3187 GB
122DeepSeek R1open412 GB · 12 benchmarksCoding 100Agents –Math –Science 99Reasoning 9512412 GB
123Kimi K2 Instruct 0905open617 GB · 6 benchmarksCoding 100Agents –Math –Science –Reasoning –6617 GB
124DeepSeek v3 0324open412 GB · 12 benchmarksCoding 100Agents –Math –Science 98Reasoning 8812412 GB
125GPT 4.1 (Apr 14, 2025)16 benchmarksCoding 99Agents –Math 92Science 96Reasoning 9216–
126Claude 3.5 Sonnet (Oct 22, 2024)10 benchmarksCoding 99Agents 100Math –Science 88Reasoning 9710–
127GLM 4.5open218 GB · 7 benchmarksCoding 99Agents 105Math –Science –Reasoning –7218 GB
128Grok 3 Mini Beta6 benchmarksCoding 99Agents –Math –Science –Reasoning –6–
129Grok Code Fast 13 benchmarksCoding 99Agents –Math –Science –Reasoning –3–
130Qwen3.5 27Bopen17.9 GB · 6 benchmarksCoding 98Agents –Math –Science –Reasoning 111617.9 GB
131DeepSeek V4 Flash Vision Expopen183 GB · 6 benchmarks—Coding 98Agents –Math –Science –Reasoning 1236183 GB
132GPT OSS 120Bopen70.8 GB · 19 benchmarksCoding 98Agents 82Math –Science 103Reasoning 961970.8 GB
133DeepSeek v3open412 GB · 8 benchmarksCoding 97Agents –Math –Science 90Reasoning 878412 GB
134GPT 4.5 Preview (Feb 27, 2025)10 benchmarksCoding 96Agents –Math –Science 98Reasoning 9610–
135GLM 4.6open217 GB · 6 benchmarksCoding 96Agents 104Math –Science 100Reasoning –6217 GB
136Qwen3 Coder Nextopen48.3 GB · 3 benchmarks—Coding 96Agents –Math –Science 97Reasoning –348.3 GB
137NVIDIA Nemotron 3 Super 120B A12B BF16open74.7 GB · 4 benchmarksCoding 96Agents –Math –Science 98Reasoning –474.7 GB
138Gemini 2.5 Flash Lite Preview Thinking (Jun 17)5 benchmarksCoding 95Agents –Math –Science –Reasoning 875–
139Mercury 2.54 benchmarksCoding 95Agents –Math –Science 99Reasoning –4–
140Trinity Large Thinkingopen271 GB · 3 benchmarks—Coding 93Agents –Math –Science 102Reasoning –3271 GB
141Qwen3 32Bopen22.6 GB · 8 benchmarksCoding 93Agents –Math –Science 96Reasoning 92822.6 GB
142Mistral Large 3 675B Instruct 2512open454 GB · 6 benchmarksCoding 93Agents –Math –Science 97Reasoning 866454 GB
143Mistral Medium 25086 benchmarksCoding 92Agents –Math –Science 91Reasoning 926–
144GPT 4.1 Mini (Apr 14, 2025)15 benchmarksCoding 92Agents –Math 95Science 97Reasoning 9115–
145O1 Mini (Sep 12, 2024)7 benchmarksCoding 91Agents –Math –Science –Reasoning 857–
146Claude 3.5 Haiku (Oct 22, 2024)8 benchmarksCoding 88Agents –Math –Science 73Reasoning –8–
147Qwen3.5 35B A3Bopen22.0 GB · 8 benchmarksCoding 86Agents –Math –Science 108Reasoning 107822.0 GB
148GPT 4o (Nov 20, 2024)10 benchmarksCoding 82Agents –Math –Science 82Reasoning 9010–
149Llama 4 Maverick 17B 128E Instructopen273 GB · 15 benchmarksCoding 81Agents –Math –Science 96Reasoning 8615273 GB
150Llama 3.3 70B Instructopen51.3 GB · 12 benchmarksCoding 76Agents –Math –Science 81Reasoning 831251.3 GB
151GPT 4.1 Nano (Apr 14, 2025)13 benchmarksCoding 75Agents –Math –Science 82Reasoning 7213–
152Gemma 3 27B ITopen21.7 GB · 9 benchmarksCoding 70Agents –Math –Science 81Reasoning 74921.7 GB
153GPT 4o Mini (Jul 18, 2024)13 benchmarksCoding 68Agents –Math 76Science –Reasoning 7313–
154Llama 3.1 8B Instructopen7.9 GB · 9 benchmarksCoding 56Agents –Math –Science 56Reasoning 7097.9 GB

The range after each score is a 90% interval: given the boards a model has results on, its true score is very likely inside it. Short ranges mean many agreeing results; long ranges mean few or conflicting ones.

Why some models are missing. A model receives a score only if it has results on at least 4 benchmarks across at least 2 categories, with one being reasoning or coding, or on 3 benchmarks that narrow its 90% range to at most 15.6 points. The results must come from at least two different benchmarks: categories of one benchmark, such as LiveBench, count once. Skill scores require at least 2 benchmarks for that skill; otherwise they display as –.

VRAM is llmrun's estimate at Q4_K_M (or the smallest quantization we track) for the weights plus room for a 16K-token context. Longer contexts need more memory. Proprietary models can't run locally, so the VRAM filter hides them.

How the llmrun Composite Score is built · llmrun doesn't run these benchmarks; scores come from public sources.

Dataset. Grab the scores of every open model as a dataset on Hugging Face, including a timestamped snapshot for each refresh.