ModelVRAM
EN 中文

What LLMs can an M1 Mac (8 GB) run?

An M1 Mac (8 GB) gives a model 5.3 GB usable of memory (about two thirds of unified memory is usable by the GPU) and 68.25 GB/s of bandwidth. Of the 61 open models tracked here, 5 fit at Q4_K_M with an 8,192-token context, and 5 more at a lower precision, each leaving at least 0.5 GB free on the M1 Mac (8 GB).

Models that fit on one M1 Mac (8 GB)

The most precise weights that still fit with 8K tokens of context and one request, the memory that takes, the longest context at that precision and the writing speed. Bigger models come first.

Model Best precision Memory Longest context Tokens/s
Ornith 1.0 9B 9.4B GGUF IQ3_XXS 4.75 GB 9K 8.9–12
MiMo V2.6 Distill Qwen 9B 9.4B GGUF IQ3_XXS 4.75 GB 9K 8.9–12
LFM2.5 8B-A1B 8.5B, 1.5B active GGUF Q2_K 4.24 GB 51K 27–46
Gemma 4 E4B 8.0B GGUF Q3_K_M 4.68 GB 14K 9.1–12
Ling 3.0 Tiny 7.9B-A1.3B 7.9B, 1.3B active GGUF Q3_K_M 4.51 GB 47K 28–48
Spark-X2.5 4B 4.1B GGUF Q6_K 4.38 GB 18K 9.8–13
Nemotron 3 Nano 4B 4.0B FP8 / INT8 4.71 GB 13K 9.0–12
Granite 4.2 3B 3.7B GGUF Q6_K 4.26 GB 14K 10–14
MiniCPM5 2B 2.5B GGUF Q8_0 3.60 GB 34K 12–17
Limite 1B Violetto 1.0B FP16 / BF16 2.79 GB 128K (full) 16–23

Too big for M1 Mac (8 GB) at 8K context

How much memory each of these lacks at Q4_K_M (gpt-oss-20b and gpt-oss-120b at MXFP4) with 8K context, one request and 0.5 GB kept free, out of the 5.3 GB usable here; a bigger machine or a smaller quant is the way in.

Best models for M1 Mac (8 GB) with room for long context

The biggest models that still fit at Q4_K_M with 32,768 tokens of context (or their whole window, if shorter), the memory left over, the longest context the Mac takes at that precision and the writing speed with that context in the cache, from 68.25 GB/s of bandwidth.

ModelContextMemoryLeft overLongest contextTokens/s
Spark-X2.5 4B 4.1B 32K 4.40 GB 919 MB 42K 9.7–13
Nemotron 3 Nano 4B 4.0B 32K 3.51 GB 1.79 GB 106K 13–17
MiniCPM5 2B 2.5B 32K 3.50 GB 1.80 GB 60K 13–17
Limite 1B Violetto 1.0B 32K 1.62 GB 3.68 GB 128K (full) 33–46

llama-server commands for an M1 Mac (8 GB)

GGUF repos checked 2026-09-29; -c is the longest context with 0.5 GB free on the M1 Mac (8 GB).

Spark-X2.5 4B, Q4_K_M with 39K tokens: 9.1–12 tokens/s on the M1 Mac (8 GB)

llama-server -hf XHToken/Spark-X2.5-4B-GGUF:Q4_K_M -c 39936 -ngl 99 -np 1

Spark-X2.5-4B-Q4_K_M.gguf, 2.6 GB: 4.79 GB used and 524 MB free of the M1 Mac (8 GB)'s 5.3 GB usable, 9.1–12 tokens/s once the 39K cache is full. -np 1: one slot, one sliding window.

Nemotron 3 Nano 4B, Q4_K_M with 77K tokens: 10–14 tokens/s on the M1 Mac (8 GB)

llama-server -hf unsloth/NVIDIA-Nemotron-3-Nano-4B-GGUF:Q4_K_M -c 78848 -ngl 99

NVIDIA-Nemotron-3-Nano-4B-Q4_K_M.gguf, 2.9 GB: 4.79 GB used and 517 MB free of the M1 Mac (8 GB)'s 5.3 GB usable, 10–14 tokens/s once the 77K cache is full.

MiniCPM5 2B, Q4_K_M with 58K tokens: 9.1–12 tokens/s on the M1 Mac (8 GB)

llama-server -hf bartowski/MiniCPM5-2B-GGUF:Q4_K_M -c 59392 -ngl 99

MiniCPM5-2B-Q4_K_M.gguf, 1.6 GB: 4.77 GB used and 541 MB free of the M1 Mac (8 GB)'s 5.3 GB usable, 9.1–12 tokens/s once the 58K cache is full.

How fast the M1 Mac (8 GB) writes as the context fills

Tokens per second at Q4_K_M for the biggest models that fit with 32K tokens, with 8K and 32K tokens in the cache and at the longest context the Mac holds, and at Q8_0 with 8K where that fits. Each token reads the active weights and the whole cache once, so the speed follows the 68.25 GB/s of bandwidth and falls as the cache grows. The ceiling is that bandwidth divided by the bytes read per token, which no runtime reaches; the reply time is for 1,000 new tokens with 32K (or the longest context) in the cache, prompt processing aside. With the same memory, M2 or M3 Mac (8 GB), 100 GB/s, writes 45% faster on average.

Model8K32KLongestAt the longestQ8_0 at 8KBandwidth ceiling at 8K1,000-token reply at 32K
Spark-X2.5 4B 13–17 9.7–13 42K 8.8–12 Does not fit 24 75–103 s
Nemotron 3 Nano 4B 14–20 13–17 106K 8.9–12 Does not fit 27 58–80 s
MiniCPM5 2B 19–27 13–17 60K 8.9–12 12–17 36 58–80 s
Limite 1B Violetto 44–63 33–46 128K 16–22 28–40 86 22–31 s

M1 Mac (8 GB) against the other Macs with 5.3 GB usable

The same models fit on every Mac with 5.3 GB usable, so speed is what separates them. By bandwidth the M1 Mac (8 GB) is the slowest of the 2 Macs with 5.3 GB usable at 68.25 GB/s. Over the 5 biggest models that fit at Q4_K_M with 8K context, the M2 or M3 Mac (8 GB), at 100 GB/s, writes 45% faster than the M1 Mac (8 GB). Each cell below is the other Mac's tokens per second minus this one's, midpoints of the estimated ranges.

Model M1 Mac (8 GB) tokens/s M2 or M3 Mac (8 GB)
Spark-X2.5 4B 13–17 +6.9
Nemotron 3 Nano 4B 14–20 +7.8
Granite 4.2 3B 13–18 +6.9
MiniCPM5 2B 19–27 +10
Limite 1B Violetto 44–63 +23

Near misses on M1 Mac (8 GB)

Models that miss at Q4_K_M with 32,768 tokens of context (or their whole window), or leave under 0.5 GB free there, but fit with a shorter context or 3-bit weights, or are short by at most a quarter of the memory and fit with MoE offload. Smallest shortfall first.

ModelShort by at Q4_K_MFits instead
Granite 4.2 3B 3.7B 224 MB Q4_K_M with 23K context
Ling 3.0 Tiny 7.9B-A1.3B 7.9B 332 MB GGUF Q3_K_M with 32K, 4.68 GB
Gemma 4 E4B 8.0B 767 MB GGUF IQ3_XXS with 32K, 4.47 GB
LFM2.5 8B-A1B 8.5B 881 MB GGUF IQ3_XXS with 32K, 4.49 GB

How these numbers are worked out

Memory is the weights at each precision plus the KV cache for 8,192 tokens in FP16 and runtime overhead (0.5 GB plus 10%), computed from each model's files on Hugging Face with the LLM VRAM Calculator. Speed is estimated from memory bandwidth and the parameters read per token with the LLM Speed Calculator. Both tools take any other model from Hugging Face.

Nearby GPUs and setups

How many of the tracked models each one holds at Q4_K_M with 8K context, against 5 here.

SetupHow it relatesMemoryBandwidthModels at Q4_K_M
M2 or M3 Mac (8 GB) Same memory 5.3 GB usable 100 GB/s 5 (same)
M1 Mac (16 GB) Next size up 10.7 GB usable 68.25 GB/s 17 (+12)

Model numbers read from Hugging Face on .