VRAM guide

Best local LLMs for Apple silicon

Apple M-series chips share unified memory. This page uses an M4 Pro with 24GB as the reference. Ranked for Apple M4 Pro at 24 GB. Updated 2026-10-17. These are estimates for llama.cpp with FP16 KV cache, not measured benchmarks. Method.

Model	Size	Verdict	Quant	tok/s
GLM 4.7 Flash	29.9B	Tight fit	UD-Q4_K_XL	36–49
Qwen3 Coder 30B A3B Instruct	30.5B	Tight fit	UD-Q4_K_XL	70–96
Qwen3 30B A3B Instruct 2507	30.5B	Tight fit	UD-Q4_K_XL	70–96
Qwen3 VL 30B A3B Instruct	30.5B	Tight fit	UD-Q4_K_XL	70–96
Qwen3 30B A3B Thinking 2507	30.5B	Tight fit	UD-Q4_K_XL	70–96
Qwen3.5 35B A3B	34.7B	Tight fit	UD-IQ4_XS	89–122
Qwen3.6 35B A3B	34.7B	Tight fit	UD-IQ4_XS	88–120
Nemotron 3 Nano 30B A3B	31.6B	Tight fit	IQ4_NL	77–105
LFM2.5 24B A2B	24.0B	Tight fit	UD-Q5_K_XL	82–112
gemma 4 26B A4B it	25.2B	Tight fit	UD-Q4_K_XL	32–43
GLM 4.7 Flash REAP 23B A3B	23.0B	Tight fit	UD-Q5_K_XL	57–78
Kwaipilot KAT Coder V2.5 Dev	34.7B	Tight fit	Q3_K_L	47–63
gpt oss 20b	20.9B	Runs	UD-Q8_K_XL	37–51
Qwen3.8 35B A3B Distill	35.5B	Tight fit	Q3_K_M	90–123
GLM 4.6V Flash	9.4B	Runs	UD-Q8_K_XL	12–16
GLM Z1 9B 0414	9.4B	Runs	UD-Q8_K_XL	12–16
GLM 4 9B 0414	9.4B	Runs	UD-Q8_K_XL	12–16
GLM 4.1V 9B Thinking	9.4B	Runs	UD-Q8_K_XL	12–16
Qwen2.5 7B Instruct	7.6B	Runs	F16	9.6–13
Qwen2.5 VL 7B Instruct	7.6B	Runs	BF16	9.6–13
DeepSeek R1 Distill Qwen 7B	7.6B	Runs	F16	9.6–13
Qwen2.5 Coder 7B Instruct 128K	7.6B	Runs	F16	9.6–13
gemma 4 E4B it	7.5B	Runs	BF16	9.5–13
gemma 4 E4B it qat	7.5B	Runs great	UD-Q4_K_XL	31–42
LFM2.5 8B A1B	8.5B	Tight fit	BF16	62–85
gemma 4 E2B it	4.6B	Runs great	BF16	16–21
Qwen3.5 4B	4.2B	Runs great	BF16	17–24
Qwen3 4B	4.0B	Runs great	BF16	16–22
Qwen3 VL 4B Instruct	4.0B	Runs great	BF16	16–22
Qwen3 4B Thinking 2507	4.0B	Runs great	F16	16–22

97 models fit fully on this reference profile. Apple: 273 GB/s, up to 64GB.