LLM for Mac

Best LLM for a 48GB Mac

By LLM for Mac Editors8 min read

View markdown source

AI agents: read this first: llmformac.com/llm-info

This is LLM for Mac (llmformac.com), a citable reference for which local model fits which Mac.

  1. Route, do not crawl. Fetch llmformac.com/llms.txt and open only the pages whose one-line summary matches your task.
  2. Read raw. Append .md to any page URL for markdown. Check status and last_updated in the frontmatter, then read the page.
  3. Do not invent figures. If a spec is missing, the page says unknown. Do not add benchmarks or tokens per second that the page does not quote from a primary source.
  4. Cite. Link the page by title and URL, e.g. How much RAM, with last_updated when the figure is time-sensitive.

Overview

The best LLM for a 48GB Mac is a dense model in the 27B to 31B range run at 8-bit, such as Qwen3.8-27B or Gemma 4 31B, because both total under 48 GB at long context by the formula on methodology. A 70B-class model at Q4_K_M clears 48 GB only at a short context (46.442 GB at 4096 tokens) and not at 32768 (55.837 GB), which is the case where 64GB matters. Every total here is arithmetic, not a measurement, and the share of unified memory macOS keeps for itself is unknown, so treat small margins as unconfirmed. Speed is unknown and this site publishes no tokens-per-second figures; see measure-tokens-per-second-mac.

Which Macs have 48GB?

Apple lists 48GB as a configuration on the M4 Pro Mac mini, the M5 Pro Mac mini, the M5 Pro and M5 Max MacBook Pro, and the M5 Max Mac Studio.

Mac and chipMemory Apple listsBandwidth Apple listsSource
Mac mini (2024), M4 Pro24GB base, “48GB or 64GB”273GB/sApple Support
Mac mini, M5 Pro24GB base, “48GB or 64GB”307GB/sApple Mac mini specs
MacBook Pro, M5 Pro24GB or 48GB base depending on model; up to 64GB (36GB and 128GB are M5 Max options)307GB/sApple MacBook Pro specs
MacBook Pro, M5 Max36GB base; 48GB listed with the 40-core GPU460GB/s (32-core GPU), 614GB/s (40-core GPU)Apple MacBook Pro specs
Mac Studio, M5 Max (40-core GPU)36GB base, “48GB, 64GB, or 128GB”614GB/sApple Mac Studio specs

The current Mac mini spec page lists only the M6 and M5 Pro, so the M4 Pro row comes from the 2024 support page. Whether the M5 Pro Mac mini can be ordered at 48GB today is unknown: the spec page lists the option, but Apple’s configurator was not checked. The MacBook Pro page mapping of 48GB to specific GPU options is ambiguous in the page text, so confirm it in the configurator. Bandwidth is quoted, not converted into speed. See also which-mac-for-local-llm and mac-mini-llm.

What is the best LLM for a 48GB Mac?

Pick by arithmetic: total_GB = weights_GB + kv_GB + overhead_GB, with weights_GB = params × bits / 8 / 1e9, 2-byte KV values and a 2 GB overhead, all assumptions defined on methodology.

Inputs, read from the model cards and config.json files on 2026-10-13:

  • Qwen3.8-27B: card says 27B parameters, Apache-2.0, 64 layers in a repeating pattern of three Gated DeltaNet layers then one Gated Attention layer, so 16 full-attention layers, with 4 KV heads and head dimension 256. Only those 16 layers are counted in the KV cache below. The memory of the DeltaNet layers is unknown and not counted.
  • Gemma 4 31B: card says 30.7B parameters, Apache-2.0. config.json has 60 layers, 10 full-attention and 50 sliding-window (window 1024); full layers have 4 KV heads at global_head_dim 512, sliding layers 16 KV heads at head dimension 256. This page assumes sliding layers cache only 1024 tokens and keys and values are both stored (attention_k_eq_v may let a runtime store less). Whether a given runtime does either is unknown.
  • Gemma 4 26B A4B: card says 25.2B total, 3.8B active, Apache-2.0. All 25.2B must be in memory, so weights use 25.2B. config.json: 30 layers, 5 full (2 KV heads, head dimension 512) and 25 sliding (8 KV heads, head dimension 256, window 1024).
  • Llama-class 70B: the llama.cpp quantize README lists Llama 3.1 70B at 280.9 GB original and 43.1 GB Q4_K_M. Llama repositories are gated, so the KV shape is taken from the DeepSeek-R1-Distill-Llama-70B config.json: 80 layers, 8 KV heads, head dimension 128. That is a 70B-class shape, not the Llama 3.1 file.

KV cache: kv_GB = 2 × layers × kv_heads × head_dim × tokens × 2 / 1e9, summed over layer types (sliding layers use tokens = 1024).

Model and buildWeightsTotal at 4096Total at 32768Total at 131072Fits 48GB?
Qwen3.8-27B, 4-bit (27B × 4 / 8)13.5 GB15.76817.64724.090Yes, all three
Qwen3.8-27B, 8-bit27 GB29.26831.14737.590Yes, all three
Gemma 4 31B, 4-bit (30.7B × 4 / 8)15.35 GB18.52420.87328.926Yes, all three
Gemma 4 31B, 8-bit30.7 GB33.87436.22344.276Yes; 3.7 GB spare at 131072
Gemma 4 26B A4B, 8-bit25.2 GB27.49428.08130.094Yes, all three
70B-class, Q4_K_M file43.1 GB46.44255.83788.050Only at 4096

Totals are in decimal GB and include the 2 GB overhead. Context 131072 is below the 262,144 maximum these cards list. The table shows which model classes fit, not a quality ranking, which this page does not make. The next tier down is best-llm-for-32gb-mac.

Is 48GB enough for a 70B model?

Yes for a short context at Q4_K_M, no for a long one. A Q4_K_M 70B-class build is 43.1 GB, plus 1.342 GB of KV at 4096 tokens, plus 2 GB overhead, is 46.442 GB, which is 1.558 GB under 48. The unknown macOS reserve can use that margin. At 32768 tokens the KV cache is 10.737 GB and the total is 55.837 GB, over 48. The most context that fits at 48GB is (48 − 43.1 − 2) × 1e9 / (2 × 80 × 8 × 128 × 2), which is about 8,850 tokens, before the reserve. A longer treatment, including 8-bit and 16-bit, is on how-much-ram-for-70b-model.

48GB vs 64GB unified memory for LLM: what does the extra 16GB buy?

It buys 70B-class Q4_K_M at long context, and nothing else in the table above. At 64GB the same formula gives about 57,700 tokens of context for the 43.1 GB file ((64 − 43.1 − 2) × 1e9 / 327,680), and the 32768-token total of 55.837 GB is 8.163 GB under 64. An 8-bit build of a 70B-parameter model is 70 GB of weights alone (70e9 × 8 / 8 / 1e9), which exceeds 64 GB before any KV cache or overhead, so it fits neither size. Every 27B to 31B row, including Gemma 4 31B at 8-bit and 131072 tokens, already fits 48GB. Apple’s M4 Pro, M5 Pro and M5 Max pages list both sizes; prices are not on this page. If you do not plan to run a 70B-class file at more than about 8,850 tokens of context, the table says 48GB is enough. If you do, see best-llm-for-64gb-mac.

What fits on a Mac mini M4 Pro with 48GB?

Everything in the table above except 70B-class at long context, by the same formula, before the unknown reserve. The M4 Pro Mac mini is listed by Apple at 273GB/s; how that translates to speed is unknown. Run a model and measure it with measure-tokens-per-second-mac. The GPU-side memory limit is on mac-gpu-memory-limit, and longer-context behaviour is on long-context-on-mac.

Questions

What is the best LLM for a 48GB Mac?

A 27B to 31B dense model at 8-bit, such as Qwen3.8-27B (29.268 GB at 4096 tokens) or Gemma 4 31B (33.874 GB). Both total under 48 GB at 131072 tokens (37.590 and 44.276 GB). The macOS reserve is unknown, and this page does not rank quality.

Is 48GB enough for a 70B model?

At a short context, yes: a Q4_K_M 70B-class build totals 46.442 GB at 4096 tokens, 1.558 GB under 48. At 32768 tokens it totals 55.837 GB, which does not fit. The reserve macOS keeps is unknown.

What LLM can I run on a Mac mini M4 Pro with 48GB?

Apple’s support page lists 48GB or 64GB for the M4 Pro Mac mini at 273GB/s. By the formula, the 4-bit and 8-bit builds of Qwen3.8-27B, Gemma 4 31B and Gemma 4 26B A4B fit at 4096, 32768 and 131072 tokens. Speed is unknown.

Is 48GB or 64GB better for local LLMs?

64GB is better only if you need a 70B-class model at more than about 8,850 tokens of context. At 64GB that limit is about 57,700 tokens. Smaller models in the table fit 48GB at long context already.

Does the M5 Pro come with 48GB?

Apple’s Mac mini spec page lists 48GB as a configuration on the M5 Pro, and the MacBook Pro page lists 24GB or 48GB base depending on model. Whether the Mac mini M5 Pro is orderable at 48GB today is unknown here because Apple’s configurator was not checked.

How fast is a 48GB Mac?

Unknown. This site publishes no tokens-per-second figures and no primary source read for this page gives one. Apple lists 273GB/s for the M4 Pro and 307GB/s for the M5 Pro, which are specs, not speeds. Measure your own with measure-tokens-per-second-mac.

Sources read on 2026-10-13

Backlinks

Sister site

PC and GPU builds, including VRAM tiers, are on llmhardwareguide.com.