Best local LLM for 8 GB.
Qwen3.5 4B is the smartest model that fits in 8 GB. It needs about 5.8 GB and is about 23% as smart as Claude Opus 5.5.
| Model | Maker | Score (AA) | Needs 4-bit | Needs 8-bit | Run it |
|---|---|---|---|---|---|
| Qwen3.5 4B | Alibaba Qwen | 13.1 | 5.8 GB | 8 GB | ollama run qwen3.5:4b |
| MiniCPM5 2B | OpenBMB | 12.5 | 3.9 GB | 5.1 GB | MLX only (Mac) |
| Granite 4.2 3B | IBM | 9.1 | 4.5 GB | 6.5 GB | ollama run granite4.2:3b |
| Qwen3.5 2B | Alibaba Qwen | 6.9 | 4.2 GB | 5.1 GB | ollama run qwen3.5:2b |
| Phi-4-mini 3.8B | Microsoft | 6.3 | 4.9 GB | 6.7 GB | ollama run phi4-mini:3.8b |
| Qwen3.5 0.8B | Alibaba Qwen | 6.1 | 2.6 GB | 3.1 GB | ollama run qwen3.5:0.8b |
Pick by what you do.
Codingnone scored
No scored coding model fits yet.
Agents (uses tools)none scored
No scored agent model fits yet.
What fits.
| Model | Needs | Score (AA) |
|---|---|---|
| Qwen3.5 4B | 5.8 GB | 13.1 |
| MiniCPM5 2B | 3.9 GB | 12.5 |
| Granite 4.2 3B | 4.5 GB | 9.1 |
| Qwen3.5 2B | 4.2 GB | 6.9 |
| Phi-4-mini 3.8B | 4.9 GB | 6.3 |
Computers and cards with 8 GB.
Graphics cards count VRAM, so they often fit a bit more.
| Machine | Type | Usable | Models that fit | Best fit | From (approx) |
|---|---|---|---|---|---|
| GeForce RTX 5060 | NVIDIA graphics card | 7 GB | 19 | Qwen3.5 4B | $370 · 2026-10-15 |
| NVIDIA Jetson Orin Nano Super Developer Kit | Dev kit | 6 GB | 19 | Qwen3.5 4B | $399 · 2026-10-12 |
Short answers.
- What is the best local LLM for 8 GB of RAM?
- Qwen3.5 4B by Alibaba Qwen. It scores 13.1 on the Artificial Analysis Intelligence Index and needs about 5.8 GB at 4-bit. That is about 23% of Claude Opus 5.5 (57.6), the top paid model. Data as of 2026-10-15.
- How many local models fit in 8 GB?
- 19 of the 124 text models in our data fit. We assume about 6 GB is usable for the model (75% of 8 GB), and a model needs its 4-bit file size × 1.15 + 2 GB.
- What is the biggest model that runs in 8 GB?
- Youtu-Parsing-Omni 5B, 5.3B parameters, needs about 5.7 GB at 4-bit. Bigger is not always smarter: Qwen3.5 4B scores 13.1; Youtu-Parsing-Omni 5B has no published score yet.
- Does 8 GB of VRAM work the same as 8 GB of RAM?
- Close. On a Mac the memory is shared, and about 75% goes to the model. On a PC the graphics card's VRAM is what counts; we estimate VRAM minus 1 GB is usable, so a 8 GB card fits a little more than 8 GB of Mac memory. Models that spill out of VRAM into normal RAM run much slower.
Scores are from Artificial Analysis and LMArena, not our own tests yet. Memory need = 4-bit file × 1.15 + 2 GB; a model fits when that is at most 75% of memory. Models 2026-10-15, machines 2026-10-15.
# Best local LLM for 8GB RAM or VRAM URL: https://insidethemachine.ai/best/local-llm-for-8gb-ram Data: models 2026-10-15, machines 2026-10-15. Source of scores: Artificial Analysis Intelligence Index (AA), LMArena. Not our own tests yet. Usable memory assumed: 6 GB (75% of 8). Rule: min_memory_gb = weights_gb_4bit × 1.15 + 2. weights_gb_4bit = Q4_K_M file (Ollama, else HF GGUF) → MLX 4-bit (HF) → Ollama MLX/NVFP4 tag → Ollama default tag → params_total_b × 0.6 (flagged weights_estimated). +15% ≈ KV cache at ~8–16K context; +2GB runtime/process overhead. Long contexts (128K+) need much more. Models that fit: 19 of 124. ## Top picks | model | maker | AA index | needs (4-bit) | needs (8-bit) | run | |---|---|---|---|---|---| | Qwen3.5 4B | Alibaba Qwen | 13.1 | 5.8 GB | 8 GB | `ollama run qwen3.5:4b` | | MiniCPM5 2B | OpenBMB | 12.5 | 3.9 GB | 5.1 GB | MLX only | | Granite 4.2 3B | IBM | 9.1 | 4.5 GB | 6.5 GB | `ollama run granite4.2:3b` | | Qwen3.5 2B | Alibaba Qwen | 6.9 | 4.2 GB | 5.1 GB | `ollama run qwen3.5:2b` | | Phi-4-mini 3.8B | Microsoft | 6.3 | 4.9 GB | 6.7 GB | `ollama run phi4-mini:3.8b` | | Qwen3.5 0.8B | Alibaba Qwen | 6.1 | 2.6 GB | 3.1 GB | `ollama run qwen3.5:0.8b` | ## By job - Coding: none scored at this size - Agents (tool use): none scored at this size - Vision (reads images): Qwen3.5 4B (5.8 GB) ## New at 8 GB - Qwen3.5 4B (5.8 GB, AA 13.1) - MiniCPM5 2B (3.9 GB, AA 12.5) - Granite 4.2 3B (4.5 GB, AA 9.1) - Qwen3.5 2B (4.2 GB, AA 6.9) - Phi-4-mini 3.8B (4.9 GB, AA 6.3) ## Machines with 8 GB | machine | type | usable GB | models that fit | best fit | from (approx) | |---|---|---|---|---|---| | GeForce RTX 5060 | NVIDIA graphics card | 7 | 19 | Qwen3.5 4B | $370 (2026-10-15) | | NVIDIA Jetson Orin Nano Super Developer Kit | Dev kit | 6 | 19 | Qwen3.5 4B | $399 (2026-10-12) | ## Questions **What is the best local LLM for 8 GB of RAM?** Qwen3.5 4B by Alibaba Qwen. It scores 13.1 on the Artificial Analysis Intelligence Index and needs about 5.8 GB at 4-bit. That is about 23% of Claude Opus 5.5 (57.6), the top paid model. Data as of 2026-10-15. **How many local models fit in 8 GB?** 19 of the 124 text models in our data fit. We assume about 6 GB is usable for the model (75% of 8 GB), and a model needs its 4-bit file size × 1.15 + 2 GB. **What is the biggest model that runs in 8 GB?** Youtu-Parsing-Omni 5B, 5.3B parameters, needs about 5.7 GB at 4-bit. Bigger is not always smarter: Qwen3.5 4B scores 13.1; Youtu-Parsing-Omni 5B has no published score yet. **Does 8 GB of VRAM work the same as 8 GB of RAM?** Close. On a Mac the memory is shared, and about 75% goes to the model. On a PC the graphics card's VRAM is what counts; we estimate VRAM minus 1 GB is usable, so a 8 GB card fits a little more than 8 GB of Mac memory. Models that spill out of VRAM into normal RAM run much slower. More: https://insidethemachine.ai/models?tier=8 · https://insidethemachine.ai/machines?tier=8 · https://insidethemachine.ai/best