# Best local LLM for 24GB RAM or VRAM

URL: https://insidethemachine.ai/best/local-llm-for-24gb-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: 18 GB (75% of 24). 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: 45 of 124.

## Top picks

| model | maker | AA index | needs (4-bit) | needs (8-bit) | run |
|---|---|---|---|---|---|
| K2 Horizon 7B | IFM (MBZUAI) | 20.6 | 8.6 GB | 13 GB | MLX only |
| Gemma 4 12B | Google | 14.2 | 11.2 GB | 16.9 GB | `ollama run gemma4:12b` |
| Qwen3.5 9B | Alibaba Qwen | 13.3 | 9.6 GB | 13.5 GB | `ollama run qwen3.5:9b` |
| 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 8B | IBM | 11.1 | 8.1 GB | 12.7 GB | `ollama run granite4.2:8b` |

## By job

- Coding: NousCoder 14B (12.3 GB) 
- Agents (tool use): Mellum2.1 12B-A2.5B Thinking (11.3 GB) 
- Vision (reads images): Gemma 4 12B (11.2 GB)

## New at 24 GB (did not fit in 16 GB)

- gpt-oss-20b (15.4 GB, AA 10)
- Ministral 3 14B (12.5 GB, AA 6)
- Phi-4 14B (12.5 GB, AA 5.9)
- Gemma 3n E4B (13.3 GB, AA 4.8)
- Phi-4-reasoning-plus 14B (14.6 GB, AA unscored)

## Machines with 24 GB

| machine | type | usable GB | models that fit | best fit | from (approx) |
|---|---|---|---|---|---|
| MacBook Air 13"/15" (M5) | Mac laptop | 18 | 45 | K2 Horizon 7B | $1,299 (2026-10-15) |
| MacBook Pro 14" (M5) | Mac laptop | 18 | 45 | K2 Horizon 7B | $1,999 (2026-10-15) |
| MacBook Pro 14"/16" (M5 Pro) | Mac laptop | 18 | 45 | K2 Horizon 7B | $2,499 (2026-10-15) |
| Mac mini (M6) | Mac desktop | 18 | 45 | K2 Horizon 7B | $899 (2026-10-15) |
| Mac mini (M5 Pro) | Mac desktop | 18 | 45 | K2 Horizon 7B | $1,699 (2026-10-15) |
| GeForce RTX 4090 | NVIDIA graphics card | 23 | 56 | Qwen3.8 27B | $3,099 (2026-10-02) |
| GeForce RTX 3090 (used) | NVIDIA graphics card | 23 | 56 | Qwen3.8 27B | $1,430 (2026-10-02) |
| GeForce RTX 3090 Ti | NVIDIA graphics card | 23 | 56 | Qwen3.8 27B | unknown |
| RTX PRO 4000 Blackwell 24GB | NVIDIA workstation card | 23 | 56 | Qwen3.8 27B | $2,900 (2026-10-15) |
| Radeon RX 7900 XTX | AMD graphics card | 23 | 56 | Qwen3.8 27B | $950 (2026-10-02) |
| Intel Arc Pro B60 (24GB) | Intel graphics card | 23 | 56 | Qwen3.8 27B | $650 (2026-10-15) |
| Tesla P40 24GB (used) | Used server card | 23 | 56 | Qwen3.8 27B | $277 (2026-10-02) |

## Questions

**What is the best local LLM for 24 GB of RAM?**
K2 Horizon 7B by IFM (MBZUAI). It scores 20.6 on the Artificial Analysis Intelligence Index and needs about 8.6 GB at 4-bit. That is about 36% of Claude Opus 5.5 (57.6), the top paid model. Data as of 2026-10-15.

**How many local models fit in 24 GB?**
45 of the 124 text models in our data fit. We assume about 18 GB is usable for the model (75% of 24 GB), and a model needs its 4-bit file size × 1.15 + 2 GB.

**What is the biggest model that runs in 24 GB?**
gpt-oss-20b, 20.9B (3.6B active) parameters, needs about 15.4 GB at 4-bit. Bigger is not always smarter: K2 Horizon 7B scores 20.6; gpt-oss-20b scores 10.

**What is the best coding model for 24 GB?**
NousCoder 14B (needs 12.3 GB). Run it in a coding app like OpenCode or Cline. See https://insidethemachine.ai/best/local-coding-model.

**Does 24 GB of VRAM work the same as 24 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 24 GB card fits a little more than 24 GB of Mac memory. Models that spill out of VRAM into normal RAM run much slower.

More: https://insidethemachine.ai/models?tier=24 · https://insidethemachine.ai/machines?tier=24 · https://insidethemachine.ai/best