Your rig

M3 · 24 GB · 24 GB

24 GB unified · 100 GB/s · 18.0 GB addressable by the GPU
Edit hardware Export as JSON

Context length

KV cache at f16, 128K tokens

Quality floor
Q8_0 — near-lossless Q6_K Q5_K_M Q4_K_M — recommended Q3_K_M Q2_K — damaged
KV cache type
Runtime
Ollamallama.cppLM StudioMLX

4 of 5 runtimes drive this device.

Allow CPU offload Include partial-offload fits

Offloaded layers run at 60 GB/s host bandwidth — expect single-digit tokens per second.

Which one should I actually run?

Best models for your M3 · 24 GB

Updated 8 Oct 2026 1 new this month

What do you want to do? Each row is the best-ranked model on that use case's shortlist that fits at Q6_K and 128K. The ordering is editorial; the sizes and speeds are computed. How the shortlists work.

RecommendationModelWhyQuant · totalTok/s est.
Best overall Qwen3.5 9B
Alibaba · Apache 2.0
The 8–12 GB default of 2026 — beats every older 8B, with vision. Q6_K · 12.0 GB 7.6 Runs great
Best for coding Qwen3.5 9B
Alibaba · Apache 2.0
The best coding model for 8–12 GB cards. Q6_K · 12.0 GB 7.6 Runs great
Best reasoning gpt-oss 20B
OpenAI · Apache 2.0
Low/medium/high reasoning effort in 13 GB. MXFP4 · 14.4 GB 14 Runs great
Best for writing Gemma 4 E4B
Google · Apache 2.0
Readable prose on a laptop. Q6_K · 8.5 GB 9.1 Runs great
Best vision Qwen3.5 9B
Alibaba · Apache 2.0
Vision on an 8–12 GB card. Q6_K · 12.0 GB 7.6 Runs great
Best for agents gpt-oss 20B
OpenAI · Apache 2.0
Harmony-format tool calling, native 4-bit. MXFP4 · 14.4 GB 14 Runs great
Best translation Qwen3.5 9B
Alibaba · Apache 2.0
Strong CJK and European translation in 8–12 GB. Q6_K · 12.0 GB 7.6 Runs great
Fastest good model gpt-oss 20B
OpenAI · Apache 2.0
3.6B active, native 4-bit. MXFP4 · 14.4 GB 14 Runs great
Best long context Qwen3.5 9B
Alibaba · Apache 2.0
262K native in 8–12 GB. Q6_K · 12.0 GB
at 128K context
7.6 Runs great

Everything that fits

Runs great

23 of 80 · fits with headroom for a longer prompt
ModelParamsQuantWeights +KV 128KTotal VRAM headroom Tok/sRuntime
gpt-oss 20B 20.9B MoE / 3.6B act MXFP4 10.8 GB 3.00 GB 14.4 GB
3.6 GB free
14 MLX Run
Phi-4 14B 14.7B Q6_K 11.2 GB 3.13 GB 15.0 GB
3.0 GB free
5.0 MLX Run
Qwen3.5 9B 9.65B Q6_K 7.4 GB 4.00 GB 12.0 GB
6.0 GB free
7.6 MLX Run
Ornith 1.5 9B 9.41B Q6_K 7.2 GB 4.00 GB 11.8 GB
6.2 GB free
7.8 MLX Run
LFM2.5 8B-A1B 8.47B MoE / 1.5B act Q6_K 6.5 GB 1.46 GB 8.5 GB
9.5 GB free
23 MLX Run
Fara 7B 8.29B Q6_K 6.3 GB 6.84 GB 13.8 GB
4.2 GB free
8.8 MLX Run
Gemma 4 E4B 8.0B Q6_K 6.1 GB 1.78 GB 8.5 GB
9.5 GB free
9.1 MLX Run
Ling 3.0 Tiny 7.9B-A1.3B 7.9B MoE / 1.3B act Q6_K 6.0 GB 0.84 GB 7.5 GB
10.5 GB free
26 MLX Run
DeepSeek-R1-Distill-Qwen 7B 7.62B Q6_K 5.8 GB 7.00 GB 13.4 GB
4.6 GB free
9.6 MLX Run
Qwen2.5-Coder 7B 7.62B Q6_K 5.8 GB 7.00 GB 13.4 GB
4.6 GB free
9.6 MLX Run
Mistral 7B Instruct v0.3 7.25B Q6_K 5.5 GB 4.00 GB 10.1 GB
7.9 GB free
10 MLX Run
Gemma 4 E2B 5.1B Q6_K 3.9 GB 0.89 GB 5.4 GB
12.6 GB free
14 MLX Run
Qwen3.5 4B 4.66B Q6_K 3.6 GB 4.00 GB 8.2 GB
9.8 GB free
16 MLX Run
Gemma 3 4B 4.3B Q6_K 3.3 GB 3.11 GB 7.0 GB
11.0 GB free
17 MLX Run
Qwen3 4B 4.02B Q6_K 3.1 GB 4.50 GB 8.2 GB
9.8 GB free
18 MLX Run
Granite 4.1 3B 3.4B Q6_K 2.6 GB 10.00 GB 13.2 GB
4.8 GB free
22 MLX Run
LFM2.5 2.6B 2.7B Q6_K 2.1 GB 2.00 GB 4.7 GB
13.3 GB free
27 MLX Run
Qwen3.5 2B 2.27B Q6_K 1.7 GB 1.50 GB 3.8 GB
14.2 GB free
32 MLX Run
Qwen3 1.7B 1.72B Q6_K 1.3 GB 3.50 GB 5.4 GB
12.6 GB free
43 MLX Run
Llama 3.2 1B Instruct 1.24B Q6_K 0.9 GB 4.00 GB 5.5 GB
12.5 GB free
59 MLX Run
Gemma 3 1B 1.0B Q6_K 0.8 GB 0.14 GB 1.5 GB
16.5 GB free
73 MLX Run
Qwen3.5 0.8B 0.87B Q6_K 0.7 GB 1.50 GB 2.8 GB
15.2 GB free
84 MLX Run
Qwen3 0.6B 0.6B Q6_K 0.5 GB 3.50 GB 4.6 GB
13.4 GB free
122 MLX Run

Runs — tight

3 models · fits at 128K, drop context or quant to go longer
ModelParamsQuantWeights +KV 128KTotal VRAM headroom Tok/sRuntime
Olmo 3 7B Instruct 7.3B Q6_K 5.6 GB 9.50 GB 15.7 GB
2.3 GB free
10 MLX Run
Ministral 3 3B 3.85B Q6_K 2.9 GB 13.00 GB 16.5 GB
1.5 GB free
19 MLX Run
Llama 3.2 3B Instruct 3.21B Q6_K 2.5 GB 14.00 GB 17.1 GB
0.9 GB free
23 MLX Run

Won't fit in VRAM

54 models · some are possible with CPU offload
ModelParamsQuantWeights +KV 128KTotal Over budget Tok/sRuntime
Kimi K3 2.8T-A104B 2780B MoE / 104B act Q6_K 2123.0 GB 3.38 GB 2127.0 GB +2109.0 GB ~0.3 MLX Why
Qwen3.8 2.4T-A95B 2446B MoE / 95B act Q6_K 1868.0 GB 11.50 GB 1880.1 GB +1862.1 GB ~0.4 MLX Why
DeepSeek V4 Pro 1.6T-A49B 1650B MoE / 49B act Q6_K 1260.1 GB 8.58 GB 1269.3 GB +1251.3 GB ~0.7 MLX Why
Kimi K2.6 1T-A32B 1027B MoE / 32B act Q6_K 784.3 GB 8.58 GB 793.5 GB +775.5 GB ~1.1 MLX Why
GLM-5.2 744B-A40B 753B MoE / 40B act Q6_K 575.1 GB 10.97 GB 586.6 GB +568.6 GB ~0.9 MLX Why
GLM-5.3 744B-A40B 753B MoE / 40B act Q6_K 575.1 GB 10.97 GB 586.6 GB +568.6 GB ~0.9 MLX Why
DeepSeek-R1 671B 671B MoE / 37B act Q6_K 512.4 GB 8.58 GB 521.6 GB +503.6 GB ~0.9 MLX Why
DeepSeek V4.1 Flash 552B-A16B New 552B MoE / 16B act Q6_K 421.6 GB 0.57 GB 422.7 GB +404.7 GB ~2.1 DwarfStar Why
Qwen3.5 397B-A17B 403B MoE / 17B act Q6_K 307.8 GB 3.75 GB 312.1 GB +294.1 GB ~2.0 MLX Why
Ornith 1.5 397B-A17B 397B MoE / 17B act Q6_K 303.2 GB 3.75 GB 307.5 GB +289.5 GB ~2.0 MLX Why
GLM-5.3-Flash 320B-A18B 321B MoE / 18B act Q6_K 245.1 GB 1.38 GB 247.1 GB +229.1 GB ~1.9 MLX Why
DeepSeek V4 Flash 284B-A13B 284B MoE / 13B act Q6_K 216.9 GB 6.05 GB 223.5 GB +205.5 GB ~2.6 MLX Why
Qwen3 235B-A22B 235B MoE / 22B act Q6_K 179.5 GB 23.50 GB 203.6 GB +185.6 GB ~1.6 MLX Why
Qwen3.8-Flash-Next 180B-A6B 180B MoE / 6B act Q6_K 137.5 GB 3.00 GB 141.1 GB +123.1 GB ~5.7 MLX Why
Qwen3.5 122B-A10B 125B MoE / 10B act Q6_K 95.5 GB 3.00 GB 99.1 GB +81.1 GB ~3.4 MLX Why
Nemotron 3 Super 120B-A12B 124B MoE / 12B act Q6_K 94.7 GB 1.00 GB 96.3 GB +78.3 GB ~2.8 MLX Why
Ling 3.0 Flash 124B-A5B 124B MoE / 5.1B act Q6_K 94.7 GB 0.98 GB 96.3 GB +78.3 GB ~6.7 MLX Why
Mistral Small 4 119B-A6B 119B MoE / 6.5B act Q6_K 90.9 GB 2.81 GB 94.3 GB +76.3 GB ~5.3 MLX Why
gpt-oss 120B 117B MoE / 5.1B act MXFP4 60.5 GB 4.50 GB 65.6 GB +47.6 GB ~9.9 MLX Why
Llama 4 Scout 109B-A17B 109B MoE / 17B act Q6_K 83.2 GB 7.13 GB 91.0 GB +73.0 GB ~2.0 MLX Why
Llama 3.3 70B Instruct 70.6B Q6_K 53.9 GB 40.00 GB 94.5 GB +76.5 GB ~1.0 MLX Why
DeepSeek-R1-Distill-Llama 70B 70.6B Q6_K 53.9 GB 40.00 GB 94.5 GB +76.5 GB ~1.0 MLX Why
Qwen3.6 35B-A3B 35.9B MoE / 3.3B act Q6_K 27.4 GB 2.50 GB 30.5 GB +12.5 GB ~10 MLX Why
Ornith 1.5 35B-A3B 35.9B MoE / 3.3B act Q6_K 27.4 GB 2.50 GB 30.5 GB +12.5 GB ~10 MLX Why
LLM-jp 4 33B Thinking 33.2B Q6_K 25.4 GB 16.00 GB 42.0 GB +24.0 GB ~2.2 MLX Why
Qwen3 32B 32.8B Q6_K 25.0 GB 32.00 GB 57.6 GB +39.6 GB ~2.2 MLX Why
DeepSeek-R1-Distill-Qwen 32B 32.8B Q6_K 25.0 GB 32.00 GB 57.6 GB +39.6 GB ~2.2 MLX Why
Qwen2.5-Coder 32B 32.8B Q6_K 25.0 GB 32.00 GB 57.6 GB +39.6 GB ~2.2 MLX Why
Olmo 3.1 32B Instruct 32.2B Q6_K 24.6 GB 4.75 GB 29.9 GB +11.9 GB ~2.3 MLX Why
Nemotron 3.5 Lightning 30B-A3B 31.6B MoE / 3.2B act Q6_K 24.1 GB 0.75 GB 25.5 GB +7.5 GB ~11 MLX Why
Gemma 4 31B 31.3B Q6_K 23.9 GB 20.78 GB 45.3 GB +27.3 GB ~2.3 MLX Why
GLM-4.7-Flash 30B-A3B 31.2B MoE / 3B act Q6_K 23.8 GB 6.61 GB 31.0 GB +13.0 GB ~11 MLX Why
Qwen3 30B-A3B 30.5B MoE / 3.3B act Q6_K 23.3 GB 12.00 GB 35.9 GB +17.9 GB ~10 MLX Why
Qwen3 Coder 30B-A3B 30.5B MoE / 3.3B act Q6_K 23.3 GB 12.00 GB 35.9 GB +17.9 GB ~10 MLX Why
Muse Glimmer 30B 29.8B Q6_K 22.8 GB 1.70 GB 25.1 GB +7.1 GB ~2.5 MLX Why
Granite 4.1 30B 28.9B Q6_K 22.1 GB 32.00 GB 54.7 GB +36.7 GB ~2.5 MLX Why
Qwen3.8 27B 27.8B Q6_K 21.2 GB 8.00 GB 29.8 GB +11.8 GB ~2.6 MLX Why
Qwen3.6 27B 27.8B Q6_K 21.2 GB 8.00 GB 29.8 GB +11.8 GB ~2.6 MLX Why
Gemma 3 27B 27.4B Q6_K 20.9 GB 10.41 GB 31.9 GB +13.9 GB ~2.7 MLX Why
Gemma 4 26B-A4B 26.5B MoE / 3.8B act Q6_K 20.2 GB 5.20 GB 26.0 GB +8.0 GB ~9.0 MLX Why
Devstral Small 2 24B 24B Q6_K 18.3 GB 20.00 GB 38.9 GB +20.9 GB ~3.0 MLX Why
Mistral Small 3.2 24B 23.6B Q6_K 18.0 GB 20.00 GB 38.6 GB +20.6 GB ~3.1 MLX Why
Qwen3 14B 14.8B Q6_K 11.3 GB 20.00 GB 31.9 GB +13.9 GB ~4.9 MLX Why
DeepSeek-R1-Distill-Qwen 14B 14.8B Q6_K 11.3 GB 24.00 GB 35.9 GB +17.9 GB ~4.9 MLX Why
Qwen2.5-Coder 14B 14.8B Q6_K 11.3 GB 24.00 GB 35.9 GB +17.9 GB ~4.9 MLX Why
Ministral 3 14B 13.9B Q6_K 10.6 GB 20.00 GB 31.2 GB +13.2 GB ~5.3 MLX Why
Gemma 3 12B 12.2B Q6_K 9.3 GB 8.31 GB 18.2 GB +0.2 GB ~6.0 MLX Why
Mistral NeMo 12B 12.2B Q6_K 9.3 GB 20.00 GB 29.9 GB +11.9 GB ~6.0 MLX Why
Gemma 4 12B 12B Q6_K 9.2 GB 8.31 GB 18.1 GB +0.1 GB ~6.1 MLX Why
Ministral 3 8B 8.92B Q6_K 6.8 GB 17.00 GB 24.4 GB +6.4 GB ~8.2 MLX Why
Granite 4.1 8B 8.79B Q6_K 6.7 GB 20.00 GB 27.3 GB +9.3 GB ~8.3 MLX Why
Qwen3 8B 8.19B Q6_K 6.3 GB 18.00 GB 24.9 GB +6.9 GB ~8.9 MLX Why
Llama 3.1 8B Instruct 8.03B Q6_K 6.1 GB 16.00 GB 22.7 GB +4.7 GB ~9.1 MLX Why
Phi-4-mini 3.8B 3.84B Q6_K 2.9 GB 16.00 GB 19.5 GB +1.5 GB ~19 MLX Why

Totals = quantised weights + f16 KV cache at 128K tokens + 0.6 GB runtime overhead, against 18.0 GB usable unified memory. Tokens per second are estimated from 100 GB/s peak memory bandwidth at batch 1 — not measured. Read the method.


Same memory, different speed

These machines have the same 18.0 GB usable, so they run exactly the models above. Generation speed scales with memory bandwidth.

DeviceBandwidthSpeed vs yours
M4 Pro · 24 GB
M4 Pro · 2024
273 GB/s 2.73×
M2 · 24 GB
M2 · 2022
100 GB/s same

Best LLM for a 24 GB Mac, with every machine in the class →

What more memory unlocks

At Q4_K_M and 8K context this machine fits 45 of 80 models. Each step up, and the largest models it adds:

Every device ranked by what it can run →