Best LLM Picks for an 8GB Mac
Computed live from the ModelFit model database for an 8GB MacBook Air: the highest-quality local models that fit inside a ~5.6GB memory budget, deduplicated so the list does not repeat the same model at a different quantization. Speed labels are ModelFit estimates, not measured benchmarks.
| # | Model | Params | Quant | Loaded Size | Speed (est.) |
|---|---|---|---|---|---|
| 1 | LFM2.5 8B-A1B | 8.3B (MoE, ~1.5B active) | Q4_K_M | ~6GB | ~71 tok/s |
| 2 | Qwen3.5 4B Instruct | 4B | Q4_K_M | ~4GB | ~79 tok/s |
| 3 | Granite 4.1 8B Instruct | 8B | Q4_K_M | ~6GB | ~34 tok/s |
| 4 | DeepSeek-R1 Distill Qwen 7B | 7B | Q4_K_M | ~6GB | ~38 tok/s |
| 5 | Gemma 4 E4B | 4.5B | Q4_K_M | ~4GB | ~71 tok/s |
Top pick: LFM2.5 8B-A1B, on-device agents, tool calling, multilingual chat.
ollama run lfm2.5:8b-a1b-q4_K_M
Params, loaded size, and speed are computed live from the ModelFit model database (data/models.json) via the same engine that powers the wizard. Speed labels are estimates, not measured results.
The 8GB RAM Budget, Explained
Apple's unified memory is shared between macOS, every open app, and the GPU that runs a local model. ModelFit sizes recommendations to a fraction of total RAM, not all of it, so there is headroom left for the operating system and whatever else is running.
Total unified memory on the machine.
ModelFit's model memory budget for this tier (70% of total RAM).
Left for macOS, background apps, and the Ollama runtime (~100MB overhead).
Inside that ~5.6GB budget, 23 models in ModelFit's catalog fit, topping out around 8.3B parameters at Q4 quantization. That ceiling is why 8GB Mac picks are small, dense models or sparse Mixture-of-Experts models rather than the larger dense models a 32GB or 64GB Mac can load.
Context Windows Cost Memory Too
Loading a model's weights is only part of the memory bill. Every token of conversation held in the context window is cached (the KV cache) so the model does not recompute it on every reply. On an 8GB Mac, that cache competes with the same small budget as the model weights.
| Context Window | Est. KV Cache | Weights + KV Cache |
|---|---|---|
| 8k tokens | ~1.0GB | ~6.5GB |
| 32k tokens | ~4.0GB | ~9.5GB |
At a 32k-token context window, LFM2.5 8B-A1B's weights plus KV cache alone exceed the ~5.6GB budget on an 8GB Mac. Keep the context window short, closer to 8k tokens, on this tier.
KV-cache figures assume an fp16 cache, the llama.cpp/Ollama default. Standard GQA models use a size-class estimate (8 KV heads x 128 head dim class); hybrid linear-attention models (Qwen3.5/3.6, Qwen3-Next) use the exact per-token cost from their published config, since only their sparse full-attention layers cache KV. A q8_0 KV cache roughly halves either figure. Estimates, not measurements.
What an 8GB Mac Cannot Run
Being honest about the ceiling matters as much as the picks that work. An 8GB Mac is not the right tier for the larger side of the local LLM catalog.
Large dense models
Dense models in the 30B-70B class need 20-45GB just to load, well past an 8GB Mac's entire RAM. These need a 32GB+ Mac at minimum.
Long, memory-heavy sessions
Long documents, large codebases, or extended chat history push the KV cache up fast on this tier. Keep sessions and context windows short.
Heavy multitasking alongside the model
A browser with many tabs, a video call, or a heavy IDE can all compete with the model for the same limited memory. Close what you can before a session.
The next tier up, for scale
16GB roughly doubles the ceiling, from about 8.3B parameters at 8GB to about 14B at 16GB.
Where to Buy for Local AI
best configsPrefer to buy direct? Buy from Apple (same price, no affiliate link).
Archive your model library off the internal drive. Quantized models run 5 to 40GB each, so 2TB holds dozens with room to spare.
Check price on Amazon40Gbps external storage fast enough to run models from. Pair it with an M.2 drive for a portable model vault.
Check price on AmazonThe fanless MacBook Air heat-soaks on long inference runs. An aluminum riser lifts the chassis so it sheds heat better off the desk.
Check price on AmazonMore ports for the external drives, displays and peripherals around a local-AI workstation.
Check price on AmazonModelFit may earn a commission on purchases through these links, at no extra cost to you.
Tips to Make 8GB Work
Use Ollama, not a heavy GUI
Ollama's server process uses only about 100MB of overhead RAM, leaving nearly all of the model budget for the model itself. See the Ollama setup guide.
Keep the quantization at Q4
Q4_K_M-class quantization is what makes an 8GB budget workable at all. Higher-precision quantizations roughly double the loaded size for a small quality gain.
Close other apps before a session
Quit browsers, video calls, and heavy editors before starting the model. Every gigabyte they hold is a gigabyte the model cannot use.
Keep the context window short
A shorter context window keeps the KV cache small. Start new conversations instead of carrying one very long session forward.
Should You Upgrade to 16GB?
Moving from 8GB to 16GB roughly doubles the comfortable model ceiling, from about 8.3B parameters to about 14B, and gives real headroom for longer context windows. At 16GB, ModelFit's top pick is Qwen3.5 9B Instruct (Q8). If local AI is an occasional convenience on an existing 8GB Mac, the picks above are genuinely usable. If it is a regular part of your workflow, more RAM is the single biggest upgrade available.
Frequently Asked Questions
Can an 8GB Mac run a large language model at all?
Yes. On an 8GB Mac, ModelFit's engine picks LFM2.5 8B-A1B (~6GB at Q4_K_M) as the highest-quality local model that fits comfortably. It is not a toy: it handles chat, summarization, and light coding help, it is just sized to a small RAM budget.
What is the best LLM for an 8GB MacBook Air?
LFM2.5 8B-A1B is the current top pick for an 8GB MacBook Air, per ModelFit's live model database. It loads to about 6GB and runs at an estimated 71 tokens per second. Pull it with `ollama run lfm2.5:8b-a1b-q4_K_M`.
How much RAM does an 8GB Mac actually have free for a model?
ModelFit budgets about 5.6GB of an 8GB Mac's unified memory for the model itself (comfortably running up to about 8.3B parameters at Q4). The rest covers macOS, background apps, and the Ollama runtime, which itself uses only about 100MB of overhead.
Does a longer context window use more memory on an 8GB Mac?
Yes. For LFM2.5 8B-A1B, ModelFit estimates the KV cache adds about 1.0GB at a 8k-token context and about 4.0GB at 32k tokens, on top of the 6GB of loaded weights. On an 8GB Mac, keep the context window short to leave headroom.
Should I upgrade to 16GB RAM for local AI on a Mac?
If you plan to use local AI regularly, yes. 16GB roughly doubles the comfortable model ceiling, from about 8.3B parameters at 8GB to about 14B at 16GB, unlocking picks like Qwen3.5 9B Instruct (Q8). See the full RAM-to-model matrix in our RAM sizing guide.