Edition 2026-11-21 · retrieved 2026-11-20 09:18 UTC
Model Leaderboards
The same scores, filtered three ways: every model, models with published weights, and the best published-weights model you can run yourself at each memory size.
Model Leaderboard
Top models by Artificial Analysis Intelligence Index
| # | Change | Model | Maker | Score |
|---|---|---|---|---|
| 1 | – | Claude Fable 5.2 (Max, Default Fallback) | Anthropic | 64.6 |
| 2 | – | GPT-6.2 Astra (Max) | OpenAI | 62.1 |
| 3 | – | Gemini 4 Neon (Xhigh) | 61.4 | |
| 4 | – | Gemini 4 Carbon (High) | 60.0 | |
| 5 | – | Claude Opus 5.5 (Max, Default Fallback) | Anthropic | 59.1 |
| 6 | – | GPT-6.2 Sol (Max) | OpenAI | 57.3 |
| 7 | – | GPT-6 Astra (Max) | OpenAI | 56.9 |
| 8 | – | GPT-6.1 Sol (Max) | OpenAI | 54.6 |
| 9 | – | Claude Sonnet 5.5 (Max, Default Fallback) | Anthropic | 54.5 |
| 10 | – | Gemini 4 Hydrogen (Xhigh) | 53.9 |
Open-Weights Leaderboard
Models whose weights the maker publishes on Hugging Face
| # | Change | Model | Maker | Score | Total params | Architecture | Memory need | License | Weights |
|---|---|---|---|---|---|---|---|---|---|
| 1 | – | DeepSeek V4.1 Flash 1017 (Max) | DeepSeek | 32.6 | 763B | ~457.9 GB | mit | deepseek-ai/DeepSeek-V4.1-Flash-1017 | |
| 2 | – | GLM-5.3 (Max) | Z AI | 31.7 | 753B | ~452 GB | other | zai-org/GLM-5.3 | |
| 3 | – | Ling 3.1 Flash 1019 | InclusionAI | 29.6 | 572B | ~343.3 GB | mit | inclusionAI/Ling-3.1-flash-1019 | |
| 4 | – | Step 5 Preview | StepFun | 28.5 | 604B | ~362.6 GB | other | stepfun-ai/Step-5-Preview | |
| 5 | – | GLM-5.3-Flash | Z AI | 27.8 | 321B | ~192.8 GB | mit | zai-org/GLM-5.3-Flash | |
| 6 | – | Muse Spark Open (Xhigh) | Meta | 24.9 | 822B | MoE, 42B active | ~493.2 GB | other | meta-models/Muse-Spark-Open-816B-A42B |
| 7 | – | DeepSeek V4 Pro 0813 (Max) | DeepSeek | 23.1 | 1.65T | ~990.3 GB | mit | deepseek-ai/DeepSeek-V4-Pro-0813 | |
| 8 | – | Qwen3.8 27B (Xhigh) | Alibaba | 20.1 | 27.8B | ~16.7 GB | apache-2.0 | Qwen/Qwen3.8-27B | |
| 9 | – | MiniMax-M3 | MiniMax | 18.9 | 427B | ~256.2 GB | other | MiniMaxAI/MiniMax-M3 | |
| 10 | – | K2 Horizon 375B A23B | Institute of Foundation Models | 17.8 | 379B | MoE, 23B active | ~227.5 GB | apache-2.0 | IFM/K2-Horizon-375B-A23B |
Best Model for 128 GB
Open-weights models whose 4-bit weights fit in 128 GB of unified memory or GPU VRAM
| # | Change | Model | Maker | Score | Total params | Architecture | Memory need | License | Weights |
|---|---|---|---|---|---|---|---|---|---|
| 1 | – | Qwen3.8 27B (Xhigh) | Alibaba | 20.1 | 27.8B | ~16.7 GB | apache-2.0 | Qwen/Qwen3.8-27B | |
| 2 | – | Nemotron 3.5 Super VL | NVIDIA | 13.6 | 124B | MoE, 12B active | ~74.7 GB | other | nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 |
| 3 | – | Muse Glimmer (High) | Meta | 6.3 | 29.8B | ~17.9 GB | apache-2.0 | meta-models/Muse-Glimmer-30B | |
| 4 | – | Nemotron 3.5 Lightning | NVIDIA | 5.5 | 31.6B | MoE, 3B active | ~18.9 GB | other | nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 |
| 5 | – | Granite 4.2 3B | IBM | 4.7 | 3.7B | ~2.2 GB | apache-2.0 | ibm-granite/granite-4.2-3b | |
| 6 | – | Mistral Small 4 (Reasoning) | Mistral | 4.3 | 119B | ~71.6 GB | apache-2.0 | mistralai/Mistral-Small-4-119B-2603 | |
| 7 | – | NVIDIA Nemotron 3 Nano 4B | NVIDIA | 1.8 | 4.0B | ~2.4 GB | other | nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 |
Best Model for 64 GB
Open-weights models whose 4-bit weights fit in 64 GB of unified memory or GPU VRAM
| # | Change | Model | Maker | Score | Total params | Architecture | Memory need | License | Weights |
|---|---|---|---|---|---|---|---|---|---|
| 1 | – | Qwen3.8 27B (Xhigh) | Alibaba | 20.1 | 27.8B | ~16.7 GB | apache-2.0 | Qwen/Qwen3.8-27B | |
| 2 | – | Muse Glimmer (High) | Meta | 6.3 | 29.8B | ~17.9 GB | apache-2.0 | meta-models/Muse-Glimmer-30B | |
| 3 | – | Nemotron 3.5 Lightning | NVIDIA | 5.5 | 31.6B | MoE, 3B active | ~18.9 GB | other | nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 |
| 4 | – | Granite 4.2 3B | IBM | 4.7 | 3.7B | ~2.2 GB | apache-2.0 | ibm-granite/granite-4.2-3b | |
| 5 | – | NVIDIA Nemotron 3 Nano 4B | NVIDIA | 1.8 | 4.0B | ~2.4 GB | other | nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 |
Best Model for 32 GB
Open-weights models whose 4-bit weights fit in 32 GB of unified memory or GPU VRAM
| # | Change | Model | Maker | Score | Total params | Architecture | Memory need | License | Weights |
|---|---|---|---|---|---|---|---|---|---|
| 1 | – | Qwen3.8 27B (Xhigh) | Alibaba | 20.1 | 27.8B | ~16.7 GB | apache-2.0 | Qwen/Qwen3.8-27B | |
| 2 | – | Muse Glimmer (High) | Meta | 6.3 | 29.8B | ~17.9 GB | apache-2.0 | meta-models/Muse-Glimmer-30B | |
| 3 | – | Nemotron 3.5 Lightning | NVIDIA | 5.5 | 31.6B | MoE, 3B active | ~18.9 GB | other | nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 |
| 4 | – | Granite 4.2 3B | IBM | 4.7 | 3.7B | ~2.2 GB | apache-2.0 | ibm-granite/granite-4.2-3b | |
| 5 | – | NVIDIA Nemotron 3 Nano 4B | NVIDIA | 1.8 | 4.0B | ~2.4 GB | other | nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 |
Best Model for 16 GB
Open-weights models whose 4-bit weights fit in 16 GB of unified memory or GPU VRAM
| # | Change | Model | Maker | Score | Total params | Architecture | Memory need | License | Weights |
|---|---|---|---|---|---|---|---|---|---|
| 1 | – | Granite 4.2 3B | IBM | 4.7 | 3.7B | ~2.2 GB | apache-2.0 | ibm-granite/granite-4.2-3b | |
| 2 | – | NVIDIA Nemotron 3 Nano 4B | NVIDIA | 1.8 | 4.0B | ~2.4 GB | other | nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 |
Scores are the Artificial Analysis Intelligence Index. Each model appears once, at its best-scoring reasoning setting. A model counts as open weights only when its maker publishes the weights on Hugging Face; total parameters and license are read from the Hugging Face API.
Memory need = total parameters × 4-bit weights × 1.2 for KV cache and runtime. A model is on a memory board when that need fits in 75% of the memory, the share macOS gives the GPU by default. On Apple Silicon and other unified-memory machines (such as AMD Strix Halo) that is total RAM; on a discrete GPU it is VRAM. Mixture-of-experts models count every expert, because all of them must be loaded; their active parameters only change speed. On a discrete GPU, experts can be offloaded to system RAM to run larger models more slowly; these boards do not assume that. Changes compare with the previous edition.