Unified memory workbench
Pick a model, quantization and context to see how much unified memory your Mac needs for a local LLM, including KV cache and headroom for macOS.
Short answer
A local LLM on a Mac needs its quantized weights plus KV cache, runtime buffers and room for macOS. As a rule of thumb at 4-bit (Q4) and 8K context: 16GB of unified memory runs 7B–8B models, 24–32GB covers 12B–32B, 64GB is the first practical tier for 70B, and 96–128GB adds room for higher precision and long context.
| Model size | Q4 weights | Minimum | Comfortable |
|---|---|---|---|
| 3B–4B | ≈ 2–3 GB | 8 GB | 16 GB |
| 7B–8B | ≈ 5 GB | 16 GB | 24 GB |
| 12B–14B | ≈ 8–9 GB | 16 GB | 24–32 GB |
| 27B–32B | ≈ 17–20 GB | 32 GB | 48 GB |
| 70B | ≈ 43 GB | 64 GB | 96–128 GB |
| 100B–120B | ≈ 60–70 GB | 96 GB | 128 GB |
| 405B | ≈ 249 GB | 512 GB | 512 GB |
Weights follow llama.cpp’s Q4_K_M examples (8B ≈ 4.9GB, 70B ≈ 43.1GB, 405B ≈ 249.1GB); minimum and comfortable include KV cache, runtime and macOS. Open the RAM guide · Last reviewed:
Model table
Computed with the same model as the calculator above: Q4_K_M, Ollama, F16 KV cache, one session. "Tight" means 8K context with 5GB for macOS and a few apps. "Comfortable" means 32K context with an 8GB reserve and at most 85% memory use. Ollama's download files are often 1–4GB larger than the pure Q4 weights because embeddings and vision towers stay at higher precision; the comfortable column leaves room for that. gpt-oss ships in MXFP4; for the 120B the Q4 estimate sits a few GB above Ollama's 65GB file, so it errs on the safe side.
| Model | Released | Q4 weights | KV cache 8K | KV cache 32K | Tight fit | Comfortable | Architecture |
|---|---|---|---|---|---|---|---|
| Qwen 3.5 9B | 5.1GB | 0.3GB | 1.0GB | 16GB | 24GB | Gated DeltaNet + attention, vision | |
| Granite 4.2 8B | 5.1GB | 1.3GB | 5.0GB | 16GB | 24GB | full attention, text only | |
| Gemma 4 12B | 6.8GB | 0.4GB | 0.6GB | 16GB | 24GB | hybrid attention, vision | |
| gpt-oss 20B (MoE) | 11.9GB | 0.2GB | 0.8GB | 24GB | 32GB | MoE, 3.6B active, text only | |
| Gemma 4 26B A4B (MoE) | 14.8GB | 0.3GB | 0.5GB | 24GB | 32GB | MoE, 4B active, all 26B resident | |
| Qwen 3.8 27B | 15.4GB | 0.6GB | 2.1GB | 24GB | 36GB | Gated DeltaNet + attention, vision | |
| Muse Glimmer 30B | 16.0GB | 0.2GB | 0.5GB | 24GB | 32GB | sliding window + global, vision | |
| Qwen3 Coder 30B A3B (MoE) | 17.4GB | 0.8GB | 3.0GB | 24GB | 36GB | MoE, 3.3B active, coding, text only | |
| Granite 4.2 30B | 16.5GB | 2.0GB | 8.0GB | 32GB | 48GB | full attention, text only | |
| Gemma 4 31B | 17.5GB | 1.1GB | 2.0GB | 32GB | 48GB | hybrid attention, vision | |
| Qwen 3.6 35B A3B (MoE) | 19.9GB | 0.2GB | 0.7GB | 32GB | 48GB | MoE, 3B active, Gated DeltaNet | |
| Llama 3.1 70B | 39.9GB | 2.5GB | 10.0GB | 64GB | 96GB | full attention | |
| gpt-oss 120B (MoE) | 66.6GB | 0.3GB | 1.1GB | 96GB | 96GB | MoE, 5.1B active, text only | |
| Qwen 3.8 Flash Next 125B A6B (MoE) | 42.2GB | 0.3GB | 0.9GB | 128GB | 128GB | MoE, 6B active, 51B n-gram table, vision | |
| GLM-5.3-Flash 320B A18B (MoE) | 182.3GB | 0.2GB | 0.7GB | 192GB | 256GB | MoE, 18B active, linear + sparse attention, vision |
The math
It fits from 32GB. It stays under 85% memory use only from 36GB; on a 24GB Mac it needs less context and a leaner reserve.
KV-Cache
Full-attention models store keys and values for every token in every layer. Gemma 4 and Muse Glimmer mostly use sliding-window layers and Qwen 3.6 and 3.8 mostly use Gated DeltaNet layers with a fixed-size state, so their cache grows far more slowly with context:
| Model | KV cache at 32K | KV cache at 128K |
|---|---|---|
| Llama 3.1 70B | 10.0GB | 40.0GB |
| Mistral Small 24B | 5.0GB | not supported |
| Gemma 4 31B | 2.0GB | 5.8GB |
| Muse Glimmer 30B | 0.5GB | 1.7GB |
| Qwen 3.8 27B | 2.1GB | 8.1GB |
In practice
Q4_K_M instead of Q6_K or Q8_0 cuts the weights by roughly a quarter to a half. Below Q4, quality losses become noticeable for most models.
Only load the context you use. With full-attention models such as Llama 3.1 70B, going from 128K to 32K saves tens of gigabytes of KV cache.
Q8_0 KV cache roughly halves cache memory with little quality loss. In Ollama it needs flash attention and OLLAMA_KV_CACHE_TYPE; llama.cpp and LM Studio have their own switches.
Gemma 4, Qwen 3.6/3.8 and Muse Glimmer keep most layers on a small fixed or windowed cache. MoE models are fast for their size, but all experts still have to sit in memory.
Browsers with many tabs, Docker and Xcode can easily hold 6–10GB. Watch memory pressure in Activity Monitor: yellow or red means the model is pushing macOS into swap.
Generation atlas
The finder covers every Apple silicon chip family in the catalog, from M1 through M6, and separates current from used options.
Workflow finder
The best Mac depends on more than the model: code, long documents, speech, and image workflows stress memory differently.
Large coding agents, repository context, and tool calls.
Balanced 24 GB+ Coding LLM (32B class)Local code models for patches, reviews, and CLI checks.
Context 24–48 GB Long documents (128K context)PDFs, source bundles, and long contexts without constant trimming.
Vision 16 GB+ Image analysisImage analysis, screenshots, OCR, and light multimodality.
Audio 8–16 GB Speech to text (Whisper)Local transcription, fast and without cloud upload.
Daily 16–24 GB All-round on the goMobile mix of chat, coding, research, and media.
FAQ
At 4-bit quantization and moderate context, plan 16GB of unified memory for 7B–8B models, 24–32GB for 12B–32B, 64GB for 70B and 96–128GB for 70B at higher precision or 100B+ models. Longer context and parallel sessions add KV cache on top.
Yes for 7B–8B models at Q4/Q5 with short to moderate context. 12B–14B models start but leave little room for a browser and IDE, so 24GB is the more comfortable entry point.
Yes. A 70B Q4_K_M file is about 43GB according to llama.cpp, so 64GB is the first practical tier with short context. 96GB or 128GB leaves room for longer context, higher-bit quantization and other apps.
Partly. On Apple silicon the CPU and GPU share one memory pool, so the GPU can address far more memory than a typical consumer graphics card. The same pool is also used by macOS and every open app, which is why the calculator keeps a reserve.
At Q4 with Ollama: 16GB runs Qwen 3.5 9B and Gemma 4 12B. 24GB runs gpt-oss 20B, and Qwen 3.8 27B (August 2026) or Meta Muse Glimmer 30B (August 2026) only at short context; 32–36GB gives those two 32K context. gpt-oss 120B needs 96GB or more. Model list checked on Hugging Face and the Ollama library on 4 October 2026.
No. Only 3B–4B parameters are active per token, which makes them fast, but every expert has to stay in memory. Plan RAM for the full 26B or 35B weights; their small KV cache is the real saving at long context.
Only small models. 1B–4B text models at Q4, such as Llama 3.2 3B or Qwen3 4B, run with short context, because macOS itself needs several gigabytes. Multimodal models like Gemma 4 E4B already ship as files of about 10GB in Ollama. For regular local AI, 16GB is the practical minimum.
Not fundamentally. All of them load quantized weights, a KV cache and some runtime buffers into the same unified memory. Differences come from the quantization format and default context, so compare at the same bit width and context length.
It is an estimate, not a benchmark. It adds model weights, KV cache, runtime buffers, optional extra parts such as vision or an n-gram table and your app reserve, and targets at most 85% memory use. Runtime version, model format and prompt length change the real figure.