Unified memory · Apple
What AI models can an Apple M2 Max 96GB run?
96 GB is server territory. The constraint stops being memory and becomes bandwidth: everything fits, and what varies is how long you wait for it. Memory is shared with the CPU, so the practical ceiling is lower than the sticker number: macOS hands roughly three quarters of it to the GPU by default.
The short answer
Assuming an 8k context window and default settings, these are the models worth downloading first.
Best all-rounder
Qwen3 Next 80B A3B Instruct
Q6_K · 63.28 GB · about 97.9 tokens/s
Best for code
Qwen3 Coder Next
Q6_K · 62.02 GB · about 97.9 tokens/s
Largest that still runs well
Qwen3.5 122B A10B
125.09B parameters · Q4_K_M · about 58.4 tokens/s
Fastest useful answer
PowerMoE 3B
about 227.6 tokens/s · 4.23 GB
See how fast it feels
Apple M2 Max 96GB running Qwen3 Next 80B A3B Instruct at Q6_K
YouWhy does my model use more memory when the conversation gets longer?
Because of the KV cache. Every token you send leaves behind a key and a value vector in each layer of the model, and those stay in memory for as long as the conversation lasts.
The weights are a fixed cost: load a 4-bit 8B model and that is about 4.8 GB, whether you write one word or ten thousand. The cache is the part that grows, and it grows in a straight line with the number of tokens in the window.
How steeply depends on the model's attention design. With grouped-query attention, several query heads share one key-value pair, which cuts the cache by that ratio. Without it, every head keeps its own, and a long context can cost more memory than the weights themselves.
If you are short on memory, the first thing to try is lowering the context window in your runtime, not the quantisation.
Reading your question
Simulated from our estimate at a 512-token question, not a recording. Assumes nothing else is competing for the GPU.
Every model, scored on this device
Each row uses the highest-quality quantisation that both fits and stays conversational. Speed is a single-stream estimate at 8k context.
| Model | Params | Verdict | Download | Memory used | Speed | Max context |
|---|---|---|---|---|---|---|
| Qwen3 Next 80B A3B Instruct Qwen | 81.32B (3.19B active) | Runs great | Q6_K | 63.28 GB | 97.9 t/s | 64k |
| Qwen3 Coder Next Qwen | 79.67B (3.19B active) | Runs great | Q6_K | 62.02 GB | 97.9 t/s | 64k |
| Qwen3.5 122B A10B Qwen | 125.09B (8.17B active) | Just fits | Q4_K_M | 71.58 GB | 58.4 t/s | 8k |
| Laguna S 2.1 Laguna | 117.56B (7.71B active) | Runs great | Q4_K_M | 66.68 GB | 70.4 t/s | 128k |
| GPT OSS 120B GPT-OSS | 116.83B (5.7B active) | Runs great | Q4_K_M | 66.18 GB | 97.1 t/s | 128k |
| GLM 4.5 Air GLM | 110.47B (13.4B active) | Runs great | Q4_K_M | 64.1 GB | 34.8 t/s | 32k |
| Mixtral 8x7B Instruct V0.1 Mistral | 46.7B (12.88B active) | Runs well | Q8_0 | 47.76 GB | 22.7 t/s | 32k |
| Phi 3.5 MoE Instruct Phi | 41.87B (6.64B active) | Runs great | Q8_0 | 42.98 GB | 41.2 t/s | 128k |
| DeepSeek V4 Flash 0731 Spark DeepSeek | 60.31B (20.36B active) | Runs well | Q8_0 | 60.24 GB | 15.5 t/s | 256k |
| Qwen3.5 35B A3B Qwen | 35.95B (2.9B active) | Runs great | Q8_0 | 36.63 GB | 89.3 t/s | 256k |
| GLM 4.7 Flash GLM | 31.22B (3.66B active) | Runs great | Q8_0 | 31.73 GB | 77.3 t/s | 128k |
| Qwen3 30B A3B Qwen | 30.53B (3.34B active) | Runs great | Q8_0 | 31.39 GB | 76.9 t/s | 32k |
| DeepSeek V4 Flash DSpark DeepSeek | 165.27B (20.36B active) | Runs great | IQ3_XXS ! | 59.44 GB | 43 t/s | 256k |
| GPT OSS 20B GPT-OSS | 20.91B (4.18B active) | Runs great | Q8_0 | 21.17 GB | 75.3 t/s | 128k |
| Gemma 4 31B Instruct Gemma | 31.27B | Runs well | Q6_K | 25.45 GB | 12.6 t/s | 256k |
| OTel 2.0 LLM 31B Instruct Other | 32.11B | Runs well | Q6_K | 26.09 GB | 12.3 t/s | 256k |
| DeepSeek Coder V2 Lite Instruct DeepSeek | 15.71B (2.74B active) | Runs great | Q8_0 | 16.21 GB | 105.8 t/s | 128k |
| Qwen3 32B Qwen | 32.76B | Runs well | Q6_K | 27.63 GB | 11.5 t/s | 32k |
| Qwen2.5 32B Instruct Qwen | 32.76B | Runs well | Q6_K | 27.63 GB | 11.5 t/s | 32k |
| Qwen3 14B Qwen | 14.77B | Runs well | Q8_0 | 16.48 GB | 19.7 t/s | 32k |
| Qwen1.5 MoE A2.7B Qwen | 14.32B (2.69B active) | Runs great | Q8_0 | 16.1 GB | 75 t/s | 8k |
| Qwen2.5 14B Instruct Qwen | 14.77B | Runs well | Q8_0 | 16.73 GB | 19.4 t/s | 32k |
| Phi 4 Phi | 14.66B | Runs well | Q8_0 | 16.68 GB | 19.4 t/s | 16k |
| Gemma 4 26B A4B Instruct Gemma | 25.81B | Runs well | Q8_0 | 26.25 GB | 12.1 t/s | 256k |
| Mistral Small 24B Instruct 2501 Mistral | 23.57B | Runs well | Q8_0 | 25.19 GB | 12.7 t/s | 32k |
| Codestral 22B V0.1 Mistral | 22.25B | Runs well | Q8_0 | 24.44 GB | 13.1 t/s | 32k |
| vLLM Translategemma 12B Instruct Gemma | 13.19B | Runs well | Q8_0 | 13.96 GB | 23.2 t/s | 128k |
| Gemma 3 27B Instruct Gemma | 27.43B | Runs well | Q8_0 | 28.86 GB | 11.1 t/s | 128k |
| Qwen3.5 27B Qwen | 27.78B | Runs well | Q8_0 | 30.1 GB | 10.6 t/s | 128k |
| Granite 4.1 30B Granite | 28.87B | Runs well | Q8_0 | 31.12 GB | 10.2 t/s | 128k |
| Mistral Nemo Instruct 2407 Mistral | 12.25B | Runs well | Q8_0 | 13.99 GB | 23.3 t/s | 128k |
| Gemma 3 12B Instruct Gemma | 12.19B | Runs well | Q8_0 | 13.41 GB | 24.2 t/s | 128k |
| Gemma 4 12B Instruct Gemma | 11.96B | Runs great | Q8_0 | 12.75 GB | 25.6 t/s | 256k |
| Llama 3.3 70B Instruct Llama | 70.55B | Fits, but slow | Q6_K | 57.18 GB | 5.5 t/s | 32k |
| Qwen2.5 72B Instruct Qwen | 72.71B | Fits, but slow | Q6_K | 58.83 GB | 5.4 t/s | 32k |
| Qwen2.5 VL 72B Instruct Qwen | 73.41B | Fits, but slow | Q6_K | 59.36 GB | 5.3 t/s | 64k |
| OLMo 2 1124 13B Instruct OLMo | 13.72B | Runs well | Q8_0 | 20.44 GB | 15.7 t/s | 4k |
| GPT NeoX 20B GPT-NeoX | 20.74B | Runs well | Q8_0 | 29.45 GB | 10.8 t/s | 2k |
| Qwen3.5 9B Qwen | 9.65B | Runs great | Q8_0 | 11.1 GB | 29.6 t/s | 256k |
| Gemma 2 9B Instruct Gemma | 9.24B | Runs great | Q8_0 | 10.98 GB | 29.8 t/s | 8k |
| Granite 4.1 8B Granite | 8.79B | Runs great | Q8_0 | 10.5 GB | 31.4 t/s | 128k |
| Fanar 1 9B Instruct Fanar | 8.78B | Runs great | Q8_0 | 10.52 GB | 31.2 t/s | 4k |
| Internlm3 8B Instruct InternLM | 8.8B | Runs great | Q8_0 | 9.63 GB | 34.4 t/s | 32k |
| LFM2.5 8B A1B Liquid | 8.47B (1.57B active) | Runs great | Q8_0 | 9.18 GB | 161.8 t/s | 64k |
| Qwen2.5 VL 7B Instruct Qwen | 8.29B | Runs great | Q8_0 | 9.16 GB | 36.1 t/s | 64k |
| DeepSeek V4 Flash DeepSeek | 290.94B (20.36B active) | Runs great | IQ2_XXS ! | 70.33 GB | 63.8 t/s | 32k |
| Qwen3 8B Qwen | 8.19B | Runs great | Q8_0 | 9.78 GB | 33.8 t/s | 32k |
| Granite 3.0 8B Instruct Granite | 8.17B | Runs great | Q8_0 | 9.88 GB | 33.4 t/s | 4k |
| T Lite Instruct 2.1 T-Lite | 8.19B | Runs great | Q8_0 | 9.78 GB | 33.8 t/s | 32k |
| Llama 3.1 8B Instruct Llama | 8.03B | Runs great | Q8_0 | 9.5 GB | 34.9 t/s | 128k |
| Apertus 8B Instruct 2509 Apertus | 8.05B | Runs great | Q8_0 | 9.52 GB | 34.8 t/s | 64k |
| Llama 3 Taiwan 8B Instruct Llama | 8.03B | Runs great | Q8_0 | 9.5 GB | 34.9 t/s | 8k |
| Gemma 4 E4B Instruct Gemma | 8B | Runs great | Q8_0 | 8.42 GB | 39.2 t/s | 128k |
| Llama 3 3 Nemotron Super 49B V1 Llama | 49.87B | Fits, but slow | Q8_0 | 70.15 GB | 4.5 t/s | 8k |
| Qwen1.5 7B Qwen | 7.72B | Runs great | Q8_0 | 12.19 GB | 26.8 t/s | 32k |
| Qwen2.5 7B Instruct Qwen | 7.62B | Runs great | Q8_0 | 8.5 GB | 39.1 t/s | 32k |
| Qwen3 235B A22B Qwen | 235.09B (22.14B active) | Runs great | IQ2_XXS ! | 58.4 GB | 46 t/s | 32k |
| MiniMax M2.7 MiniMax | 228.69B (10.98B active) | Runs great | IQ2_XXS ! | 57.27 GB | 68.3 t/s | 32k |
| MiniMax M2.5 MiniMax | 228.7B (10.98B active) | Runs great | IQ2_XXS ! | 57.27 GB | 68.3 t/s | 32k |
| OLMo 3 7B Instruct OLMo | 7.3B | Runs great | Q8_0 | 9.77 GB | 33.8 t/s | 64k |
| Command A Plus 05 2026 Bf16 Command | 218.75B (18.52B active) | Runs great | IQ2_XXS ! | 53.51 GB | 63.1 t/s | 128k |
| Mistral 7B Instruct V0.3 Mistral | 7.25B | Runs great | Q8_0 | 8.72 GB | 38.2 t/s | 32k |
| Mistral 7B Instruct V0.2 Mistral | 7.24B | Runs great | Q8_0 | 8.71 GB | 38.2 t/s | 32k |
| Falcon 7B Falcon | 7.22B | Runs great | Q8_0 | 12.16 GB | 26.9 t/s | 8k |
| DeepSeek Coder 7B Instruct V1.5 DeepSeek | 6.91B | Runs great | Q8_0 | 11.14 GB | 29.5 t/s | 4k |
| OLMoE 1B 7B 0125 Instruct OLMo | 6.92B (1.28B active) | Runs great | Q8_0 | 8.27 GB | 137.7 t/s | 4k |
| CodeLlama 7B Llama | 6.74B | Runs great | Q8_0 | 11.22 GB | 29.2 t/s | 16k |
| DeepSeek Coder 6.7B Instruct DeepSeek | 6.74B | Runs great | Q8_0 | 11.22 GB | 29.2 t/s | 16k |
| Gemma 4 E2B Instruct Gemma | 5.12B | Runs great | Q8_0 | 5.48 GB | 61.4 t/s | 128k |
| Qwen3.5 4B Qwen | 4.66B | Runs great | Q8_0 | 6.07 GB | 55.6 t/s | 256k |
| Agents A1 4B Other | 4.54B | Runs great | Q8_0 | 5.95 GB | 56.8 t/s | 256k |
| Gemma 3 4B Instruct Gemma | 4.3B | Runs great | Q8_0 | 5.01 GB | 68.5 t/s | 128k |
| Phi 3 Vision 128k Instruct Phi | 4.15B | Runs great | Q8_0 | 7.59 GB | 43.9 t/s | 128k |
| Qwen3 4B Qwen | 4.02B | Runs great | Q8_0 | 5.56 GB | 61.1 t/s | 32k |
| Phi 4 Mini Instruct Phi | 3.84B | Runs great | Q8_0 | 5.29 GB | 65 t/s | 128k |
| Phi 3 Mini 4k Instruct Phi | 3.82B | Runs great | Q8_0 | 5.02 GB | 68.9 t/s | 4k |
| PowerLM 3B PowerLM | 3.51B | Runs great | Q8_0 | 6.73 GB | 49.6 t/s | 4k |
| Granite 4.1 3B Granite | 3.4B | Runs great | Q8_0 | 4.45 GB | 78.2 t/s | 128k |
| PowerMoE 3B PowerLM | 3.37B (0.88B active) | Runs great | Q8_0 | 4.23 GB | 227.6 t/s | 4k |
| Llama 3.2 3B Instruct Llama | 3.21B | Runs great | Q8_0 | 4.54 GB | 77 t/s | 128k |
| Qwen2.5 3B Instruct Qwen | 3.09B | Runs great | Q8_0 | 3.77 GB | 93.4 t/s | 32k |
| SmolLM3 3B Base SmolLM | 3.08B | Runs great | Q8_0 | 4.04 GB | 86.4 t/s | 64k |
| Starcoder2 3B StarCoder | 3.03B | Runs great | Q8_0 | 3.6 GB | 100.1 t/s | 16k |
| Phi 2 Phi | 2.78B | Runs great | Q8_0 | 5.71 GB | 59.4 t/s | 2k |
| Step 3.5 Flash Step | 199.38B | Fits, but slow | IQ2_XXS ! | 49.07 GB | 6.4 t/s | 256k |
| LFM2.5 2.6B Liquid | 2.7B | Runs great | Q8_0 | 3.57 GB | 99.3 t/s | 128k |
| Gemma 2 2B Instruct Gemma | 2.61B | Runs great | Q8_0 | 3.43 GB | 104.4 t/s | 8k |
| Qwen3.5 2B Qwen | 2.27B | Runs great | Q8_0 | 3.05 GB | 119 t/s | 256k |
| OneRec 1.7B Other | 2.13B | Runs great | Q8_0 | 3.41 GB | 104.6 t/s | 32k |
| Qwen3 1.7B Qwen | 2.03B | Runs great | Q8_0 | 3.31 GB | 108.2 t/s | 32k |
Showing the 90 best results of 133. The remaining 43 need more memory than this device has, at any quantisation.
What it will not run
16 of the 133 architectures we track are out of reach here, even at two-bit precision.
If you want more headroom
The next steps up in memory, in order. More memory changes which models load at all; more bandwidth changes how fast they answer, so the two are worth weighing separately.
| Device | Memory | Bandwidth | Models that fit | |
|---|---|---|---|---|
| NVIDIA DGX Spark 128GB | 128 GB | 273 GB/s | 121 | Check price |
| Ryzen AI Max+ 395 128GB | 128 GB | 256 GB/s | 121 | Check price |
Price links go to an Amazon search for the model name and are affiliate links: if you buy through one, we earn a commission at no cost to you. We do not take payment for placement, and the ordering above is by memory capacity alone.
Other devices with 96 GB
Same capacity, different speed. Once a model fits, bandwidth is what separates these.