Best local LLM for 32 GB.

# Best local LLM for 32GB RAM or VRAM

URL: https://insidethemachine.ai/best/local-llm-for-32gb-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: 24 GB (75% of 32). 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: 62 of 124.

## Top picks

| model | maker | AA index | needs (4-bit) | needs (8-bit) | run |
|---|---|---|---|---|---|
| Qwen3.8 27B | Alibaba Qwen | 33.7 | 22.7 GB | 36.5 GB | `ollama run qwen3.8:27b` |
| Qwen3.5 27B | Alibaba Qwen | 22.9 | 21.5 GB | 36.5 GB | `ollama run qwen3.5:27b` |
| Qwen3.6 27B | Alibaba Qwen | 21.4 | 21.5 GB | 36.5 GB | `ollama run qwen3.6:27b` |
| K2 Horizon 7B | IFM (MBZUAI) | 20.6 | 8.6 GB | 13 GB | MLX only |
| Muse Glimmer 30B | Meta | 17.5 | 22.7 GB | 37.6 GB | `ollama run muse-glimmer:30b` |
| Gemma 4 26B-A4B | Google | 16.7 | 22.7 GB | 34.2 GB | `ollama run gemma4:26b` |

## By job

- Coding: Qwen3.8 27B (22.7 GB, Terminal-Bench 2.1 79.8%) `ollama run qwen3.8:27b`
- Agents (tool use): Qwen3.8 27B (22.7 GB) `ollama run qwen3.8:27b`
- Vision (reads images): Qwen3.8 27B (22.7 GB)

## New at 32 GB (did not fit in 24 GB)

- Qwen3.8 27B (22.7 GB, AA 33.7)
- Qwen3.5 27B (21.5 GB, AA 22.9)
- Qwen3.6 27B (21.5 GB, AA 21.4)
- Muse Glimmer 30B (22.7 GB, AA 17.5)
- Gemma 4 26B-A4B (22.7 GB, AA 16.7)

## Machines with 32 GB

| machine | type | usable GB | models that fit | best fit | from (approx) |
|---|---|---|---|---|---|
| MacBook Air 13"/15" (M5) | Mac laptop | 24 | 62 | Qwen3.8 27B | $1,299 (2026-10-15) |
| MacBook Pro 14" (M5) | Mac laptop | 24 | 62 | Qwen3.8 27B | $1,999 (2026-10-15) |
| Mac mini (M6) | Mac desktop | 24 | 62 | Qwen3.8 27B | $899 (2026-10-15) |
| GeForce RTX 5090 | NVIDIA graphics card | 31 | 75 | Qwen3.8 27B | $5,700 (2026-10-15) |
| RTX PRO 4500 Blackwell 32GB | NVIDIA workstation card | 31 | 75 | Qwen3.8 27B | $4,500 (2026-10-15) |
| Radeon AI PRO R9700 (32GB) | AMD graphics card | 31 | 75 | Qwen3.8 27B | $1,600 (2026-10-15) |
| Intel Arc Pro B70 (32GB) | Intel graphics card | 31 | 75 | Qwen3.8 27B | $1,280 (2026-10-15) |
| AMD Instinct MI50 32GB (used) | Used server card | 31 | 75 | Qwen3.8 27B | $599 (2026-10-02) |
| Tesla V100 PCIe 32GB (used) | Used server card | 31 | 75 | Qwen3.8 27B | $670 (2026-10-02) |
| Sipeed MaixStation | Dev kit | 24 | 62 | Qwen3.8 27B | unknown |

## Questions

**What is the best local LLM for 32 GB of RAM?**
Qwen3.8 27B by Alibaba Qwen. It scores 33.7 on the Artificial Analysis Intelligence Index and needs about 22.7 GB at 4-bit. That is about 59% of Claude Opus 5.5 (57.6), the top paid model. Data as of 2026-10-15.

**How many local models fit in 32 GB?**
62 of the 124 text models in our data fit. We assume about 24 GB is usable for the model (75% of 32 GB), and a model needs its 4-bit file size × 1.15 + 2 GB.

**What is the biggest model that runs in 32 GB?**
OLMo 3.1 32B Think, 32.2B parameters, needs about 23.8 GB at 4-bit. Bigger is not always smarter: Qwen3.8 27B scores 33.7; OLMo 3.1 32B Think scores 7.1.

**What is the best coding model for 32 GB?**
Qwen3.8 27B (needs 22.7 GB, Terminal-Bench 2.1 79.8%). Run it in a coding app like OpenCode or Cline. See https://insidethemachine.ai/best/local-coding-model.

**Does 32 GB of VRAM work the same as 32 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 32 GB card fits a little more than 32 GB of Mac memory. Models that spill out of VRAM into normal RAM run much slower.

More: https://insidethemachine.ai/models?tier=32 · https://insidethemachine.ai/machines?tier=32 · https://insidethemachine.ai/best