CanItRun Logocanitrun.

Apple M3

The Apple M3 ships in 8–24 GB unified-memory configurations at 100 GB/s. Across those configurations it runs 47 of our 99 tracked models natively in VRAM at 8k context.

More memory means more of our tracked models fit natively; see which configuration you need below.

Jump to:24 GB16 GB8 GB
ConfigurationBandwidthCPU coresNative models+ Offload
24 GB100 GB/s8 (4P + 4E)47 / 990
16 GB100 GB/s8 (4P + 4E)28 / 990
8 GB100 GB/s8 (4P + 4E)0 / 990
Vendor: Apple
Memory type: LPDDR5
Compute backend: METAL
Software: MLX gives the best performance on Apple Silicon; llama.cpp Metal backend is a solid alternative. Both are well-supported by Ollama.

How much of the Apple M3's memory is actually usable?

macOS and background apps need a slice of the pool before a model gets to use it: this site reserves 8GB on every unified-memory GPU, the same baseline used everywhere else on this site. What's left is real headroom for a model's weights and KV cache:

macOS + background apps (8 GB reserved)usable for model weights + KV cacheVertical lines mark roughly where a 7B/14B/32B/... dense model at Q4_K_M lands.

Apple M3 (24GB)

With 24 GB LPDDR5 at 100 GB/s, this configuration runs 47 models natively. With 16 GB of real headroom, this configuration is where mixture-of-experts models start to separate from same-size dense ones: Qwen 3.5 35B-A3B, a 35B-parameter model, fits at Q2_K (15.12 GB) at 20.1 tok/s, while the dense Qwen3 32B needs the same Q2_K precision at a similar footprint (15.50 GB) but manages only 5.8 tok/s, a 3.5x gap driven entirely by MoE decode reading far fewer bytes per token. 47 of the 99 tracked models fit natively, though most 27B+ dense models are squeezed into Q2_K or Q3_K_M here, real capability, not much precision headroom.

Apple M3 (24GB): the top memory tier of the base M3 chip, a BTO-only upgrade over the 8GB and 16GB standard configurations at the same 8-core CPU (4P+4E) and 100 GB/s bandwidth.

With 16 GB of real headroom, this configuration is where mixture-of-experts models start to separate from same-size dense ones: Qwen 3.5 35B-A3B, a 35B-parameter model, fits at Q2_K (15.12 GB) at 20.1 tok/s, while the dense Qwen3 32B needs the same Q2_K precision at a similar footprint (15.50 GB) but manages only 5.8 tok/s, a 3.5x gap driven entirely by MoE decode reading far fewer bytes per token. 47 of the 99 tracked models fit natively, though most 27B+ dense models are squeezed into Q2_K or Q3_K_M here, real capability, not much precision headroom.

MLX and llama.cpp's Metal backend are both mature on this chip. This configuration's largest fits (~15-16 GB) stay comfortably under the default ~18 GB Metal working-set limit (75% of 24 GB), so the manual wired_limit override isn't necessary here.

Models the 24 GB configuration runs natively (47)

Show 42 more

Apple M3 (16GB)

With 16 GB LPDDR5 at 100 GB/s, this configuration runs 28 models natively. With 8 GB of real headroom after the standard 8 GB reservation, this site's calculator fits Llama 3.1 8B at Q5_K_M (7.58 GB, 11.8 tok/s) and Qwen2.5 7B at Q6_K (7.51 GB, 11.9 tok/s), about 1.5x the identically-sized base M1 16GB's speed on the same two models (8 and 8.1 tok/s there), tracking the 100-vs-68-GB/s bandwidth ratio closely. 28 of the 99 tracked models fit natively, the same count as the base M1 and M2 16GB configurations at this identical 8 GB headroom, since capacity, not bandwidth, decides the model list.

Apple M3 (16GB): the 16GB build of the base M3 chip, same 8-core CPU (4P+4E) and 100 GB/s bandwidth as the 8GB and 24GB configurations; only the memory pool changes. 100 GB/s is a 47% jump over the base M1's 68 GB/s, but flat against the base M2's own 100 GB/s, the only tier in the M3 generation where bandwidth didn't move at all from the chip it replaced.

With 8 GB of real headroom after the standard 8 GB reservation, this site's calculator fits Llama 3.1 8B at Q5_K_M (7.58 GB, 11.8 tok/s) and Qwen2.5 7B at Q6_K (7.51 GB, 11.9 tok/s), about 1.5x the identically-sized base M1 16GB's speed on the same two models (8 and 8.1 tok/s there), tracking the 100-vs-68-GB/s bandwidth ratio closely. 28 of the 99 tracked models fit natively, the same count as the base M1 and M2 16GB configurations at this identical 8 GB headroom, since capacity, not bandwidth, decides the model list.

MLX and llama.cpp's Metal backend are both mature on this chip. Nothing this configuration fits comes close to the default ~12 GB Metal working-set limit (75% of 16 GB), the largest fit here is under 8 GB, so the manual wired_limit override other Apple Silicon pages need doesn't apply at this capacity.

Models the 16 GB configuration runs natively (28)

Show 23 more

Apple M3 (8GB)

With 8 GB LPDDR5 at 100 GB/s, this configuration runs 0 models natively. This site reserves 8 GB of unified memory for macOS and background apps on every Apple Silicon GPU (the same baseline used everywhere on this site), which leaves this 8 GB machine with 0 GB of real headroom for a model's weights and KV cache. None of the 99 models tracked on this site fit natively in memory here; even the smallest quantized models need more room than the reservation leaves free.

Apple M3 (8GB): the entry tier of the November 2023 M3 generation: 8-core CPU (4P+4E), 8- or 10-core GPU depending on SKU, and 100 GB/s LPDDR5 bandwidth, identical to the base M2's bandwidth two years earlier. Ships in the base 14-inch MacBook Pro, iMac, and, from March 2024, the 13-inch and 15-inch MacBook Air.

This site reserves 8 GB of unified memory for macOS and background apps on every Apple Silicon GPU (the same baseline used everywhere on this site), which leaves this 8 GB machine with 0 GB of real headroom for a model's weights and KV cache. None of the 99 models tracked on this site fit natively in memory here; even the smallest quantized models need more room than the reservation leaves free.

MLX and llama.cpp's Metal backend both run on this chip, but neither changes the arithmetic: an 8 GB pool minus an 8 GB OS reservation is 0 GB free, regardless of software. Anyone shopping specifically for local LLM inference should look at the 16GB M3 at minimum, or an M3 Pro for real headroom.

Models the 8 GB configuration runs natively (0)

None.

Too large for any Apple M3 configuration (52)

Explore

Browse all guides →Browse all models →Browse all GPUs →

Frequently asked questions

How much memory does the Apple M3 have?
The Apple M3 ships in 3 unified-memory configurations: 24 GB, 16 GB, 8 GB, all at 100 GB/s.
Should I get the 8 GB or 24 GB Apple M3?
The 8 GB configuration doesn't run any of the models tracked on this site natively in VRAM: after this site's standard OS-memory reservation, it has no real headroom left for a model's weights. The 24 GB configuration is the practical minimum for local LLM inference on this chip, fitting Qwen 3.5 35B-A3B (MoE), Ornith 1.5 35B-A3B (MoE), Qwen3 32B, and 44 more models natively.
How much VRAM does the Apple M3 (24GB) have?
The Apple M3 (24GB) has 24 GB of LPDDR5 with 100 GB/s memory bandwidth (unified system memory, shared between CPU and GPU).
What is the Apple M3 (24GB) best for?
With 24 GB of VRAM, the Apple M3 (24GB) is well-suited for running 7B–32B models at Q4 with room for context, making it a great all-rounder for local LLM inference.
What LLMs can the Apple M3 (24GB) run locally?
The Apple M3 (24GB) can run 47 of the 99 open-weight models tracked by CanItRun natively in VRAM at 8k context. Top options include: Qwen 3.8 27B at Q3_K_M, Ornith 1.5 35B-A3B (MoE) at Q2_K, Ornith 1.5 9B at Q8_0.
Can the Apple M3 (24GB) run Gemma 4 31B?
Yes. The Apple M3 (24GB) runs Gemma 4 31B natively in VRAM at Q2_K quantization, achieving approximately 6.1 tokens per second.
Show 14 more questions
Can the Apple M3 (24GB) run Qwen 3.6 27B?
Yes. The Apple M3 (24GB) runs Qwen 3.6 27B natively in VRAM at Q3_K_M quantization, achieving approximately 5.9 tokens per second.
Can the Apple M3 (24GB) run Qwen3 8B?
Yes. The Apple M3 (24GB) runs Qwen3 8B natively in VRAM at Q8_0 quantization, achieving approximately 8.2 tokens per second.
How much VRAM does the Apple M3 (16GB) have?
The Apple M3 (16GB) has 16 GB of LPDDR5 with 100 GB/s memory bandwidth (unified system memory, shared between CPU and GPU).
What is the Apple M3 (16GB) best for?
With 16 GB of VRAM, the Apple M3 (16GB) handles smaller models (7B–14B) at Q4–Q5 quantization, ideal for entry-level local LLM experimentation and lightweight inference.
What LLMs can the Apple M3 (16GB) run locally?
The Apple M3 (16GB) can run 28 of the 99 open-weight models tracked by CanItRun natively in VRAM at 8k context. Top options include: Ornith 1.5 9B at Q5_K_M, Qwen 3.5 9B at Q5_K_M, Bonsai 27B at 1-bit (Q1_0).
Can the Apple M3 (16GB) run Gemma 4 31B?
The Apple M3 (16GB) does not have enough VRAM to run Gemma 4 31B. You would need more VRAM or a lower quantization level.
Can the Apple M3 (16GB) run Qwen 3.6 27B?
The Apple M3 (16GB) does not have enough VRAM to run Qwen 3.6 27B. You would need more VRAM or a lower quantization level.
Can the Apple M3 (16GB) run Qwen3 8B?
Yes. The Apple M3 (16GB) runs Qwen3 8B natively in VRAM at Q5_K_M quantization, achieving approximately 11.6 tokens per second.
How much VRAM does the Apple M3 (8GB) have?
The Apple M3 (8GB) has 8 GB of LPDDR5 with 100 GB/s memory bandwidth (unified system memory, shared between CPU and GPU).
What is the Apple M3 (8GB) best for?
With 8 GB of VRAM, the Apple M3 (8GB) is best for running compact models (1B–8B) at low quantization, suitable for edge inference, prototyping, and lightweight tasks.
What LLMs can the Apple M3 (8GB) run locally?
The Apple M3 (8GB) cannot run any of the 99 tracked models fully in VRAM at 8k context. It may handle smaller models with CPU offload.
Can the Apple M3 (8GB) run Gemma 4 31B?
The Apple M3 (8GB) does not have enough VRAM to run Gemma 4 31B. You would need more VRAM or a lower quantization level.
Can the Apple M3 (8GB) run Qwen 3.6 27B?
The Apple M3 (8GB) does not have enough VRAM to run Qwen 3.6 27B. You would need more VRAM or a lower quantization level.
Can the Apple M3 (8GB) run Qwen3 8B?
The Apple M3 (8GB) does not have enough VRAM to run Qwen3 8B. You would need more VRAM or a lower quantization level.