MacBook Neo A18 Pro 8GB — local LLM fit & speed

Apple · 8 GB LPDDR5X · 60 GB/s · 20 W · MSRP $0

METAL

The MacBook Neo A18 Pro 8GB is a Apple accelerator with 8 GB LPDDR5X and about 60 GB/s peak memory bandwidth. AI Hashrate estimates local LLM decode (batch 1) at 8K context: fit counts weights + KV cache + 1 GB runtime overhead — not weights alone. At Q4, about 11 curated models fit fully on this card; at FP16, about 2 fit. Top Q4 speeds: MiniCPM5-1B ≈ 70.0 tok/s (estimated, fits); Qwen3.5-2B ≈ 37.8 tok/s (estimated, fits); Llama-3.2-3B-Instruct ≈ 23.6 tok/s (estimated, fits); Nanbeige4.2-3B ≈ 19.2 tok/s (estimated, fits); Qwen3-4B-Instruct-2507 ≈ 18.9 tok/s (estimated, fits). Relative ranking is more reliable than absolute tok/s. See methodology for the bandwidth formula and measured-anchor policy. Methodology.

🛒 Where to Buy

Search Amazon ↗

affiliate link · we earn from qualifying purchases

Table: decode speed estimates (Q4 / FP16). Measured rows override estimates. MSRP is list price, not live retail. Context default 8K.

Model	Params (B)	Quant	tok/s	Fits?
MiniCPM5-1B	1.08	Q4	70.0 est.	Yes
Qwen3.5-2B	2.0	Q4	37.8 est.	Yes
Llama-3.2-3B-Instruct	3.2	Q4	23.6 est.	Yes
MiniCPM5-1B	1.08	FP16	19.6 est.	Yes
Nanbeige4.2-3B	4.0	Q4	19.2 est.	Yes
Qwen3-4B-Instruct-2507	4.0	Q4	18.9 est.	Yes
NVIDIA-Nemotron-3-Nano-4B-BF16	4.0	Q4	18.9 est.	Yes
Qwen3-4B-Thinking-2507	4.0	Q4	18.9 est.	Yes
gemma-4-E2B-it	5.1	Q4	14.8 est.	Yes
Qwen3.5-9B	8.95	Q4	11.2 est.	No
Ornith-1.0-9B	9.0	Q4	11.0 est.	No
Qwen3.5-2B	2.0	FP16	10.6 est.	Yes
Qwen3-8B	8.2	Q4	10.5 est.	No
Mistral-7B-Instruct-v0.3	7.2	Q4	10.5 est.	Tight
DeepSeek-R1-0528-Qwen3-8B	8.2	Q4	10.5 est.	No
Qwen2.5-7B-Instruct	7.6	Q4	9.9 est.	Tight
Llama-3.1-8B-Instruct	8.0	Q4	9.6 est.	No
gemma-4-E4B-it	8.0	Q4	9.4 est.	Tight
gpt-oss-20b	21.5	Q4	5.7 est.	No
Llama-3.2-3B-Instruct	3.2	FP16	5.1 est.	No
gemma-4-12b-it	11.95	Q4	4.7 est.	No
gemma-4-26B-A4B-it	25.2	Q4	3.9 est.	No
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16	30.0	Q4	3.5 est.	No
Qwen3-30B-A3B-Instruct-2507	30.5	Q4	3.3 est.	No
Qwen3.5-35B-A3B	34.7	Q4	2.8 est.	No
Ornith-1.0-35B	35.0	Q4	2.7 est.	No
Qwen3.6-35B-A3B	36.0	Q4	2.6 est.	No
Nanbeige4.2-3B	4.0	FP16	2.6 est.	No
Qwen3-4B-Instruct-2507	4.0	FP16	2.5 est.	No
GLM-4.7-Flash	31.2	Q4	2.5 est.	No
Qwen3-4B-Thinking-2507	4.0	FP16	2.5 est.	No
NVIDIA-Nemotron-3-Nano-4B-BF16	4.0	FP16	2.4 est.	No
Qwen3-14B	14.8	Q4	2.2 est.	No
Qwen2.5-14B-Instruct	14.8	Q4	2.1 est.	No
DeepSeek-R1-Distill-Qwen-14B	14.8	Q4	2.1 est.	No
gemma-4-E2B-it	5.1	FP16	1.6 est.	No
Devstral-Small-2-24B-Instruct-2512	24.0	Q4	0.6 est.	No
Qwen2.5-7B-Instruct	7.6	FP16	0.5 est.	No
Mistral-7B-Instruct-v0.3	7.2	FP16	0.5 est.	No
Qwen3.5-27B	27.0	Q4	0.5 est.	No
Qwen3-8B	8.2	FP16	0.4 est.	No
Llama-3.1-8B-Instruct	8.0	FP16	0.4 est.	No
Qwen3.6-27B	27.8	Q4	0.4 est.	No
gemma-4-E4B-it	8.0	FP16	0.4 est.	No
DeepSeek-R1-0528-Qwen3-8B	8.2	FP16	0.4 est.	No
Qwen3.5-9B	8.95	FP16	0.3 est.	No
gemma-4-31B-it	30.7	Q4	0.3 est.	No
gemma-3-27b-it	27.4	Q4	0.3 est.	No
Ornith-1.0-9B	9.0	FP16	0.3 est.	No
Qwen2.5-32B-Instruct	32.8	Q4	0.2 est.	No
Qwen2.5-Coder-32B-Instruct	32.8	Q4	0.2 est.	No
DeepSeek-R1-Distill-Qwen-32B	32.8	Q4	0.2 est.	No

Fit checks

Qwen3-8B — does not fit · ≈10.5 tok/s Q4 est.
gemma-4-26B-A4B-it — does not fit · ≈3.9 tok/s Q4 est.
Qwen2.5-7B-Instruct — Tight · ≈9.9 tok/s Q4 est.
Qwen3.5-9B — does not fit · ≈11.2 tok/s Q4 est.
Llama-3.1-8B-Instruct — does not fit · ≈9.6 tok/s Q4 est.
gpt-oss-20b — does not fit · ≈5.7 tok/s Q4 est.
Qwen3.6-35B-A3B — does not fit · ≈2.6 tok/s Q4 est.
Qwen3.5-35B-A3B — does not fit · ≈2.8 tok/s Q4 est.
gemma-4-E2B-it — fits · ≈14.8 tok/s Q4 est.
Qwen3.6-27B — does not fit · ≈0.4 tok/s Q4 est.
Mistral-7B-Instruct-v0.3 — Tight · ≈10.5 tok/s Q4 est.

All 67 model fit checks — see the linked Fits? column in the table above