Best local LLM for your memory.
Memory decides which AI models fit. 16 GB is the useful start: 38 models fit, best is K2 Horizon 7B. Find your size below.
| Memory | Usable | Models that fit | Best model | vs Claude Opus 5.5 | Best for coding |
|---|---|---|---|---|---|
| 8 GB | 6 GB | 19 | Qwen3.5 4B | 23% | — |
| 16 GB | 12 GB | 38 | K2 Horizon 7B | 36% | Ornith 1.5 9B |
| 24 GB | 18 GB | 45 | K2 Horizon 7B | 36% | NousCoder 14B |
| 32 GB | 24 GB | 62 | Qwen3.8 27B | 59% | Qwen3.8 27B |
| 48 GB | 36 GB | 76 | Qwen3.8 27B | 59% | Qwen3.8 27B |
| 64 GB | 48 GB | 77 | Qwen3.8 27B | 59% | Qwen3.8 27B |
| 96 GB | 72 GB | 83 | Qwen3.8 27B | 59% | Qwen3.8 27B |
| 128 GB | 96 GB | 90 | Qwen3.8 27B | 59% | Qwen3.8 27B |
| 192 GB | 144 GB | 94 | Qwen3.8-Flash-Next (125B-A6B) | 69% | Qwen3.8-Flash-Next (125B-A6B) |
| 256 GB | 192 GB | 95 | Qwen3.8-Flash-Next (125B-A6B) | 69% | Qwen3.8-Flash-Next (125B-A6B) |
| 512 GB | 384 GB | 102 | GLM-5.3-Flash | 73% | GLM-5.3-Flash |
More guides.
Short answers.
- How much RAM do I need to run a local LLM?
- 8 GB runs small models (best: Qwen3.5 4B). 16 GB is the useful minimum: 38 models fit and the best is K2 Horizon 7B (36% of Claude Opus 5.5). 32 GB fits 62 models including Qwen3.8 27B (59%). Data as of 2026-10-15.
- Is 64 GB of RAM enough for a local LLM?
- Yes for most uses. 77 of 124 text models fit in 64 GB. The smartest is Qwen3.8 27B (AA 33.7). More memory mainly adds room for long context and bigger mixture-of-experts models.
- What can 128 GB of RAM run?
- 90 of 124 models fit. The top pick is Qwen3.8 27B (AA 33.7, about 59% of Claude Opus 5.5). You can also run it at 8-bit for better quality.
- Does more RAM make the model smarter?
- Only up to a point. In our data the best model that fits is the same (Qwen3.8 27B) from 32 GB to 128 GB. In that range more memory buys longer context, 8-bit quality and room for other apps.
Scores come from Artificial Analysis and LMArena, not our own tests yet. Memory need = 4-bit file × 1.15 + 2 GB. Data 2026-10-15.
# Best local LLM by memory (RAM or VRAM) URL: https://insidethemachine.ai/best Data: 2026-10-15. Scores: Artificial Analysis Intelligence Index. Top paid model for reference: Claude Opus 5.5 = 57.6. | memory GB | usable GB | models that fit | best model | AA | best coding | page | |---|---|---|---|---|---|---| | 8 | 6 | 19 | Qwen3.5 4B | 13.1 | - | https://insidethemachine.ai/best/local-llm-for-8gb-ram | | 16 | 12 | 38 | K2 Horizon 7B | 20.6 | Ornith 1.5 9B | https://insidethemachine.ai/best/local-llm-for-16gb-ram | | 24 | 18 | 45 | K2 Horizon 7B | 20.6 | NousCoder 14B | https://insidethemachine.ai/best/local-llm-for-24gb-ram | | 32 | 24 | 62 | Qwen3.8 27B | 33.7 | Qwen3.8 27B | https://insidethemachine.ai/best/local-llm-for-32gb-ram | | 48 | 36 | 76 | Qwen3.8 27B | 33.7 | Qwen3.8 27B | https://insidethemachine.ai/best/local-llm-for-48gb-ram | | 64 | 48 | 77 | Qwen3.8 27B | 33.7 | Qwen3.8 27B | https://insidethemachine.ai/best/local-llm-for-64gb-ram | | 96 | 72 | 83 | Qwen3.8 27B | 33.7 | Qwen3.8 27B | https://insidethemachine.ai/best/local-llm-for-96gb-ram | | 128 | 96 | 90 | Qwen3.8 27B | 33.7 | Qwen3.8 27B | https://insidethemachine.ai/best/local-llm-for-128gb-ram | | 192 | 144 | 94 | Qwen3.8-Flash-Next (125B-A6B) | 39.8 | Qwen3.8-Flash-Next (125B-A6B) | https://insidethemachine.ai/best/local-llm-for-192gb-ram | | 256 | 192 | 95 | Qwen3.8-Flash-Next (125B-A6B) | 39.8 | Qwen3.8-Flash-Next (125B-A6B) | https://insidethemachine.ai/best/local-llm-for-256gb-ram | | 512 | 384 | 102 | GLM-5.3-Flash | 41.8 | GLM-5.3-Flash | https://insidethemachine.ai/best/local-llm-for-512gb-ram | Other guides: https://insidethemachine.ai/best/local-coding-model · https://insidethemachine.ai/best/local-model-for-agents · https://insidethemachine.ai/best/local-llm-for-mac · https://insidethemachine.ai/best/local-llm-for-nvidia-gpu · https://insidethemachine.ai/best/local-llm-for-mini-pc ## Questions **How much RAM do I need to run a local LLM?** 8 GB runs small models (best: Qwen3.5 4B). 16 GB is the useful minimum: 38 models fit and the best is K2 Horizon 7B (36% of Claude Opus 5.5). 32 GB fits 62 models including Qwen3.8 27B (59%). Data as of 2026-10-15. **Is 64 GB of RAM enough for a local LLM?** Yes for most uses. 77 of 124 text models fit in 64 GB. The smartest is Qwen3.8 27B (AA 33.7). More memory mainly adds room for long context and bigger mixture-of-experts models. **What can 128 GB of RAM run?** 90 of 124 models fit. The top pick is Qwen3.8 27B (AA 33.7, about 59% of Claude Opus 5.5). You can also run it at 8-bit for better quality. **Does more RAM make the model smarter?** Only up to a point. In our data the best model that fits is the same (Qwen3.8 27B) from 32 GB to 128 GB. In that range more memory buys longer context, 8-bit quality and room for other apps.