The Local AI Leaderboard
Artificial Analysis ranks raw frontier intelligence - and almost all of it is proprietary, cloud-only, and metered by the token. We flip it: this board ranks the best models you can actually self-host, defaulting to open weights and surfacing the VRAM and params you need to run them. Same Intelligence / Coding / Math / Agentic indexes - reframed for $0/mo local instead of pay-per-call cloud.
Best you can run locally
MiMo-V2.6-Pro leads the open field at 46.3 Intelligence - ~92 GB VRAM to run. Buy the GPU once, then $0/mo forever. No tokens, no rate limits, your data stays home.
Frontier, but cloud-only
Claude Opus 5.5 tops the proprietary board at 57.6 - but you can't download it. It's metered per token, can change or vanish overnight, and every prompt leaves your machine. Use the toggle below to compare it against what you can self-host.
Rankings by task
Above, one index to rule them all. Below, official per-task scores so you can shop for the job you actually have: shipping code, running agents, answering hard questions, math, or reading documents. Every number is an official leaderboard result, re-fetched on every data refresh.
Coding
Writing and shipping real code - SWE-bench scores come from models resolving GitHub issues end-to-end in real repos.
| # | Model | Params | sweVerified | swePro |
|---|---|---|---|---|
| 1 | Qwen/Qwen3.5-397B-A17B | 403.4B | 76.4 | — |
| 2 | MiniMaxAI/MiniMax-M2.5 | 228.7B | 75.8 | — |
| 3 | stepfun-ai/Step-3.5-Flash | 199.4B | 74.4 | — |
| 4 | zai-org/GLM-5 | 753.9B | 72.8 | — |
| 5 | Qwen/Qwen3.5-27B | 27.8B | 72.4 | — |
| 6 | Qwen/Qwen3.5-122B-A10B | 125.1B | 72 | — |
| 7 | moonshotai/Kimi-K2.5 | 1058.6B | 70.8 | — |
| 8 | Qwen/Qwen3-Coder-Next | 79.7B | 70.6 | 44.3 |
| 9 | deepseek-ai/DeepSeek-V3.2 | 685.4B | 70 | 15.56 |
| 10 | Qwen/Qwen3.5-35B-A3B | 36B | 69.2 | — |
Agentic use
Multi-step work with tools: navigating a terminal, running commands, and finishing supervised-free tasks.
| # | Model | Params | terminalBench |
|---|---|---|---|
| 1 | Qwen/Qwen3.5-397B-A17B | 403.4B | 52.5 |
| 2 | zai-org/GLM-5 | 753.9B | 52.4 |
| 3 | stepfun-ai/Step-3.5-Flash | 199.4B | 51 |
| 4 | Qwen/Qwen3.5-122B-A10B | 125.1B | 49.4 |
| 5 | moonshotai/Kimi-K2.5 | 1058.6B | 43.2 |
| 6 | Qwen/Qwen3.5-27B | 27.8B | 41.6 |
| 7 | Qwen/Qwen3.5-35B-A3B | 36B | 40.5 |
| 8 | deepseek-ai/DeepSeek-V3.2 | 685.4B | 39.6 |
| 9 | Qwen/Qwen3-Coder-Next | 79.7B | 36.2 |
| 10 | moonshotai/Kimi-K2-Thinking | — | 35.7 |
Knowledge & reasoning
Exam-style reasoning - graduate-level science (GPQA), broad professional knowledge (MMLU-Pro), frontier hard questions (HLE).
| # | Model | Params | mmluPro | gpqa | hle |
|---|---|---|---|---|---|
| 1 | MiniMaxAI/MiniMax-M2.1 | 228.7B | 88 | — | 22.2 |
| 2 | Qwen/Qwen3.5-397B-A17B | 403.4B | 87.8 | 88.4 | 28.7 |
| 3 | moonshotai/Kimi-K2.5 | 1058.6B | 87.1 | 87.6 | 50.2 |
| 4 | Qwen/Qwen3.5-122B-A10B | 125.1B | 86.7 | 86.6 | 25.3 |
| 5 | Qwen/Qwen3.5-27B | 27.8B | 86.1 | 85.5 | 24.3 |
| 6 | Qwen/Qwen3.5-35B-A3B | 36B | 85.3 | 84.2 | 22.4 |
| 7 | deepseek-ai/DeepSeek-R1-0528 | 684.5B | 85 | — | — |
| 8 | deepseek-ai/DeepSeek-V3.2 | 685.4B | 85 | 82.4 | 40.8 |
| 9 | moonshotai/Kimi-K2-Thinking | — | 84.6 | 84.5 | 23.9 |
| 10 | Qwen/Qwen3-235B-A22B-Thinking-2507 | 235B | 84.4 | — | — |
Math
From grade-school word problems (GSM8K) to competition rounds (AIME, HMMT).
| # | Model | Params | gsm8k | aime2026 | hmmt2026 |
|---|---|---|---|---|---|
| 1 | Qwen/Qwen2-72B | 72.7B | 89.5 | — | — |
| 2 | deepseek-ai/DeepSeek-V3 | — | 89.3 | — | — |
| 3 | microsoft/Phi-3.5-mini-instruct | — | 86.2 | — | — |
| 4 | meta-llama/Llama-3.1-8B-Instruct | 8B | 84.5 | — | — |
| 5 | moonshotai/Kimi-K2.5 | 1058.6B | — | 95.83 | 87.12 |
| 6 | Qwen/Qwen3.5-397B-A17B | 403.4B | — | 93.33 | 87.88 |
| 7 | stepfun-ai/Step-3.5-Flash | 199.4B | — | 96.67 | 86.36 |
| 8 | zai-org/GLM-5 | 753.9B | — | 95.83 | 86.36 |
| 9 | deepseek-ai/DeepSeek-V3.2 | 685.4B | — | 94.17 | 84.09 |
Vision & documents
Reading real-world documents and screenshots (olmOCR-bench).
| # | Model | Params | olmOcr |
|---|---|---|---|
| 1 | rednote-dots-ocr-community/dots.ocr-1.5 | 3B | 83.9 |
| 2 | lightonai/LightOnOCR-2-1B | 1B | 83.2 |
| 3 | datalab-to/chandra | 8.8B | 83.1 |
| 4 | infly/Infinity-Parser-7B | 8.3B | 82.5 |
| 5 | allenai/olmOCR-2-7B-1025-FP8 | 8.3B | 82.4 |
| 6 | PaddlePaddle/PaddleOCR-VL | 1B | 80 |
| 7 | rednote-hilab/dots.ocr | — | 79.1 |
| 8 | deepseek-ai/DeepSeek-OCR-2 | — | 76.3 |
| 9 | lightonai/LightOnOCR-1B-1025 | 1B | 76.1 |
| 10 | deepseek-ai/DeepSeek-OCR | 3.3B | 75.7 |
Official scores from the Hugging Face OpenEvals leaderboards - updated 2026-10-19. Models we cover link to their local page with VRAM math and run instructions; the rest are listed for comparison.
Found a model that fits your GPU? Each one has a full page with benchmarks, VRAM math, and recommended hardware.
Browse all local models