What can you run with 128GB of RAM?
With 128GB RAM you can run 30 of the 51 open-source models we track — up to about 226B parameters at Q4 quantization. The most popular that fits is Qwen3.8-27B, needing 19 GB and generating around ~22 tok/s on typical hardware of this size.
Typical machines: Mac Studio M3 Ultra · Threadripper workstation · Multi-GPU server. After the operating system takes its share, about 133 GB is free for a model.
Having the memory is not the same as having the speed
128GB RAM decides whether a model loads. How fast it answers is decided by memory bandwidth — how quickly your machine can read the model, which it must do once for every word it writes. Machines with 128GB RAM span 90–819 GB/s, so the fastest is about 9.1× quicker than the slowest with identical capacity. The table quotes a M4 Max; pick yours below to see the difference. Why bandwidth and not the processor →
Models that run on 128GB RAM in 2026
| Model↕ | Size↕ | Memory↑ | Est. speed↕ | Downloads↕ | Feels like |
|---|---|---|---|---|---|
| LFM2.5-230MLiquid Ai | 230M | 808 MBRuns well | ~2313 tok/s | 58,011 | Fast |
| MiniCPM5-1BOpenBMB | 1.1B | 1.3 GBRuns well | ~544 tok/s | 667,154 | Fast |
| MiniCPM5-2BOpenBMB | 2.5B | 2.3 GBRuns well | ~227 tok/s | 206,774 | Fast |
| LFM2.5-2.6BLiquid Ai | 2.7B | 2.5 GBRuns well | ~212 tok/s | 98,713 | Fast |
| LFM2.5-VL-3BLiquid Ai | 3.1B | 2.5 GBRuns well | ~212 tok/s | 31,296 | Fast |
| North-Micro-Vision-InstructCohere | 2.5B | 2.5 GBRuns well | ~237 tok/s | 120,354 | Fast |
| granite-4.2-3bIbm | 3.7B | 3.1 GBRuns well | ~158 tok/s | 32,017 | Fast |
| Mage-VLMicrosoft | 4.7B | 3.9 GBRuns well | ~131 tok/s | 29,690 | Fast |
| LFM2.5-8B-A1BLiquid Ai · MoE | 8.5B | 5.9 GBRuns well | ~583 tok/s1B active | 80,632 | Fast |
| Ling-3.0-tinyInclusionai · MoE | 7.9B | 6.2 GBRuns well | ~368 tok/s1.6B active | 25,727 | Fast |
| granite-4.2-8bIbm | 8.8B | 6.6 GBRuns well | ~66 tok/s | 52,946 | Fast |
| Ornith-1.5-9BDeepreinforce | 9.7B | 6.9 GBRuns well | ~61 tok/s | 453,339 | Fast |
| gemma-4-12B-itGoogle DeepMind | 12B | 9.6 GBRuns well | ~49 tok/s | 2,958,440 | Fast |
| Qwen3.8-27BAlibaba / Qwen | 28B | 19 GBRuns well | ~22 tok/s | 7,703,400 | Comfortable |
| diffusiongemma-26B-A4B-itGoogle DeepMind · MoE | 26B | 19 GBRuns well | ~136 tok/s4B active | 708,393 | Fast |
| Muse-Glimmer-30BMeta AI | 30B | 20 GBRuns well | ~20 tok/s | 441,881 | Comfortable |
| granite-4.2-30bIbm | 29B | 21 GBRuns well | ~20 tok/s | 17,894 | Comfortable |
| North-Mini-Code-1.0Cohere · MoE | 30B | 22 GBRuns well | ~92 tok/s6.1B active | 11,742 | Fast |
| Laguna-XS-2.1Poolside · MoE | 33B | 23 GBRuns well | ~88 tok/s6.7B active | 32,445 | Fast |
| Nex-N2.5-miniNex Agi · MoE | 35B | 24 GBRuns well | ~84 tok/s7B active | 4,543 | Fast |
| Ornith-1.5-35B-A3BDeepreinforce · MoE | 36B | 24 GBRuns well | ~196 tok/s3B active | 359,808 | Fast |
| NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4NVIDIA · MoE | 18B | 28 GBRuns well | ~84 tok/s3B active | 1,261,605 | Fast |
| NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4NVIDIA · MoE | 45B | 63 GBRuns well | ~33 tok/s9B active | 6,549 | Fast |
| MiniMax-H3MiniMax | 33B | 64 GBRuns well | ~6.5 tok/s | 4,827,156 | Slow |
| Laguna-M.1Poolside · MoE | 226B | 76 GBRuns well | ~25 tok/s45B active | 4,037 | Comfortable |
| Ling-3.0-flash-FinInclusionai · MoE | 127B | 84 GBRuns well | ~23 tok/s25B active | 1,268 | Comfortable |
| Ling-3.0-flashInclusionai · MoE | 127B | 85 GBRuns well | ~22 tok/s25B active | 16,875 | Comfortable |
| Ling-3.0-flash-VLInclusionai · MoE | 125B | 85 GBRuns well | ~22 tok/s25B active | 3,769 | Comfortable |
| Laguna-S-2.1Poolside · MoE | 118B | 100 GBRuns well | ~18 tok/s24B active | 17,496 | Comfortable |
| Ling-2.6-flash-baseInclusionai · MoE | 108B | 114 GBTight | ~16 tok/s22B active | 503 | Comfortable |
Sorted by memory, smallest first — click any column to change it.
All figures at Q4_K_M quantization, 4K context. “Memory” includes the model, its context cache and runtime overhead. Speed is calculated, not benchmarked: 546 GB/s × 0.65 ÷ bytes read per token, assuming a M4 Max. Scale it by your own machine's bandwidth from the list above. Mixture-of-experts models read only their active parameters per token, which is why some large ones outrun smaller dense models.
What you'd need more memory for
The next models up, and what they ask for.