Best LLM for 8 GB unified memory · 2026
Best LLM for a 8 GB Mac: what fits, and what to run
macOS lets the GPU wire down about 5.4 GB of a 8 GB Mac's unified memory; the rest stays with the system and your apps. At Q4_K_M with an 8K context, 19 of 80 models in the catalogue fit — 15 with headroom, 4 tightly. The pick for most people is Qwen3.5 4B: The 6–8 GB pick: a real assistant in 3.4 GB of weights. Speeds are estimated for an M2 · 8 GB; the fit verdicts are the same for every machine in the class.
Which one should I run?
One pick per job, from the editorial shortlist, checked to fit at Q4_K_M and 8K. Each label links the full ranking for that job. How picks are chosen.
| Recommendation | Model | Why | Quant · total | Tok/s est. | |
|---|---|---|---|---|---|
| Best overall | Qwen3.5 4B Alibaba · Apache 2.0 |
The 6–8 GB pick: a real assistant in 3.4 GB of weights. | Q4_K_M · 3.5 GB | 21 | Runs great |
| Best for coding | Qwen3.5 4B Alibaba · Apache 2.0 |
A coding agent in 3.4 GB — the 6–8 GB answer. | Q4_K_M · 3.5 GB | 21 | Runs great |
| Best reasoning | Qwen3.5 4B Alibaba · Apache 2.0 |
Step-by-step reasoning in 8 GB. | Q4_K_M · 3.5 GB | 21 | Runs great |
| Best for writing | Qwen3.5 4B Alibaba · Apache 2.0 |
Competent drafting in 8 GB. | Q4_K_M · 3.5 GB | 21 | Runs great |
| Best vision | Qwen3.5 4B Alibaba · Apache 2.0 |
Reads screenshots in 8 GB. | Q4_K_M · 3.5 GB | 21 | Runs great |
| Best for agents | Qwen3.5 4B Alibaba · Apache 2.0 |
Function calling in 8 GB. | Q4_K_M · 3.5 GB | 21 | Runs great |
| Best translation | Qwen3.5 4B Alibaba · Apache 2.0 |
Usable translation in 8 GB. | Q4_K_M · 3.5 GB | 21 | Runs great |
| Fastest good model | Ling 3.0 Tiny 7.9B-A1.3B inclusionAI · MIT |
1.3B active with a latent cache on 6 of 24 layers. | Q4_K_M · 5.1 GB | 36 | Tight fit |
| Best long context | Qwen3.5 2B Alibaba · Apache 2.0 |
Long documents on 4 GB. | Q4_K_M · 3.4 GB at 128K context |
44 | Runs great |
The long-context row is judged at 128K: Qwen3.5 2B holds that window on a 8 GB Mac at Q4_K_M with 1.50 GB of KV cache — that is why it can differ from the best overall pick.
Everything that fits a 8 GB Mac
Largest first. "Max context" is the longest window the Mac holds at that quantisation with an f16 cache; a q8_0 cache roughly doubles it.
| Model | Params | Quant | Weights | +KV 8K | Total | Headroom | Max context | Tok/s est. | |
|---|---|---|---|---|---|---|---|---|---|
| Gemma 4 E2B | 5.1B | Q4_K_M | 2.9 GB | 0.07 GB | 3.5 GB | 1.9 GB | 128K | 19 | Runs great |
| Qwen3.5 4B | 4.66B | Q4_K_M | 2.6 GB | 0.25 GB | 3.5 GB | 1.9 GB | 69K | 21 | Runs great |
| Gemma 3 4B | 4.3B | Q4_K_M | 2.4 GB | 0.30 GB | 3.3 GB | 2.1 GB | 96K | 23 | Runs great |
| Qwen3 4B | 4.02B | Q4_K_M | 2.3 GB | 1.13 GB | 4.0 GB | 1.4 GB | 18K | 25 | Runs great |
| Ministral 3 3B | 3.85B | Q4_K_M | 2.2 GB | 0.81 GB | 3.6 GB | 1.8 GB | 25K | 26 | Runs great |
| Phi-4-mini 3.8B | 3.84B | Q4_K_M | 2.2 GB | 1.00 GB | 3.8 GB | 1.6 GB | 21K | 26 | Runs great |
| Granite 4.1 3B | 3.4B | Q4_K_M | 1.9 GB | 0.63 GB | 3.1 GB | 2.3 GB | 36K | 29 | Runs great |
| Llama 3.2 3B Instruct | 3.21B | Q4_K_M | 1.8 GB | 0.88 GB | 3.3 GB | 2.1 GB | 27K | 31 | Runs great |
| LFM2.5 2.6B | 2.7B | Q4_K_M | 1.5 GB | 0.13 GB | 2.2 GB | 3.2 GB | 128K | 37 | Runs great |
| Qwen3.5 2B | 2.27B | Q4_K_M | 1.3 GB | 0.09 GB | 2.0 GB | 3.4 GB | 256K | 44 | Runs great |
| Qwen3 1.7B | 1.72B | Q4_K_M | 1.0 GB | 0.88 GB | 2.4 GB | 3.0 GB | 32K | 58 | Runs great |
| Llama 3.2 1B Instruct | 1.24B | Q4_K_M | 0.7 GB | 0.25 GB | 1.5 GB | 3.9 GB | 128K | 80 | Runs great |
| Gemma 3 1B | 1.0B | Q4_K_M | 0.6 GB | 0.04 GB | 1.2 GB | 4.2 GB | 32K | 99 | Runs great |
| Qwen3.5 0.8B | 0.87B | Q4_K_M | 0.5 GB | 0.09 GB | 1.2 GB | 4.2 GB | 256K | 114 | Runs great |
| Qwen3 0.6B | 0.6B | Q4_K_M | 0.3 GB | 0.88 GB | 1.8 GB | 3.6 GB | 32K | 166 | Runs great |
| Gemma 4 E4B | 8.0B | Q4_K_M | 4.5 GB | 0.14 GB | 5.2 GB | 0.2 GB | 19K | 12 | Tight |
| Ling 3.0 Tiny 7.9B-A1.3B | 7.9B MoE | Q4_K_M | 4.4 GB | 0.05 GB | 5.1 GB | 0.3 GB | 54K | 36 | Tight |
| DeepSeek-R1-Distill-Qwen 7B | 7.62B | Q4_K_M | 4.3 GB | 0.44 GB | 5.3 GB | 0.1 GB | 9K | 13 | Tight |
| Qwen2.5-Coder 7B | 7.62B | Q4_K_M | 4.3 GB | 0.44 GB | 5.3 GB | 0.1 GB | 9K | 13 | Tight |
Just out of reach
The smallest models that do not fit. Unified memory has no slower pool to spill into, so there is no offload fallback: each needs a bigger machine, or a lower quantisation.
| Model | Params | Total @ Q4_K_M | Over budget | Needs |
|---|---|---|---|---|
| LFM2.5 8B-A1B | 8.47B MoE | 5.5 GB | +0.1 GB | 16 GB Mac |
| Mistral 7B Instruct v0.3 | 7.25B | 5.7 GB | +0.3 GB | 16 GB Mac |
| Fara 7B | 8.29B | 5.7 GB | +0.3 GB | 16 GB Mac |
| Llama 3.1 8B Instruct | 8.03B | 6.1 GB | +0.7 GB | 16 GB Mac |
| Ornith 1.5 9B | 9.41B | 6.1 GB | +0.7 GB | 16 GB Mac |
| Qwen3.5 9B | 9.65B | 6.3 GB | +0.9 GB | 16 GB Mac |
Running it on a Mac
The GPU's share can be raised. The cap macOS puts on wired GPU memory is a setting, not a hardware limit: sudo sysctl iogpu.wired_limit_mb=… lifts it until the next reboot. Leave several gigabytes for macOS itself — every number on this page assumes the default share, so it under-promises rather than over.
MLX or GGUF. On Apple Silicon, MLX (via mlx-lm or LM Studio) is usually faster than llama.cpp's GGUF on the same model, and the community converts most popular releases within days. GGUF through Ollama or llama.cpp has the widest coverage and usually supports a new architecture first. The memory arithmetic is the same either way.
Speed is bandwidth. Every machine in this class holds the same models; what separates them is memory bandwidth (M2 100 GB/s, M3 100 GB/s, M1 68 GB/s), and tokens per second scales with it.
Every total is quantised weights + f16 KV cache at 8K + 0.6 GB runtime overhead, against 5.4 GB usable. Tokens per second are estimated from memory bandwidth and never measured. The method.