Phi · 32B to 80B · Mixture of experts
Phi 3.5 MoE Instruct
41.87 billion parameters in total, but only 6.64 billion are read for each token. That split is the whole point of the design: it costs the memory of a large model and the speed of a small one. Grouped-query attention keeps the cache small, so long contexts cost less here than on models of the same size.
Memory needed, by quantisation
At an 8k context window. Quality is our rough ranking of how much the compression costs you: anything at or above Q5 is hard to tell apart from the original in normal use.
| Quantisation | File size | KV cache | Total VRAM | Quality | What it costs you |
|---|---|---|---|---|---|
| Q8_0 | 41.43 GB | 1 GB | 43.28 GB | 99% | Lossless in practice. Use it when the memory is there. |
| Q6_K | 31.98 GB | 1 GB | 33.83 GB | 98% | Very close to Q8 for two thirds of the size. |
| Q5_K_M | 27.73 GB | 1 GB | 29.59 GB | 96% | The quality-first choice when Q6 will not fit. |
| Q4_K_M | 23.54 GB | 1 GB | 25.39 GB | 93% | The default. Best size-to-quality ratio for local use. |
| IQ4_XS | 20.72 GB | 1 GB | 22.57 GB | 91% | Importance-matrix 4-bit. Q4_K_S quality, smaller file. |
| Q3_K_M | 19.06 GB | 1 GB | 20.91 GB | 86% | Degradation starts to show. A way to fit one size up. |
| IQ3_XXS | 14.92 GB | 1 GB | 16.77 GB | 79% | Aggressive. Only worth it on very large models. |
What a longer conversation costs
Same model at Q4_K_M, only the context window changes. This model uses sliding-window attention, so most layers stop growing past 128k tokens and the bill flattens out.
| Context | KV cache | Total VRAM | Fits in 8 GB | Fits in 12 GB | Fits in 24 GB |
|---|---|---|---|---|---|
| 4k | 0.5 GB | 24.77 GB | no | no | no |
| 8k | 1 GB | 25.39 GB | no | no | no |
| 16k | 2 GB | 26.64 GB | no | no | no |
| 32k | 4 GB | 29.14 GB | no | no | no |
| 64k | 8 GB | 34.14 GB | no | no | no |
| 128k | 16 GB | 44.14 GB | no | no | no |
The "fits" columns allow for the roughly 0.8 GB Windows keeps for the desktop.
Which hardware runs Phi 3.5 MoE Instruct
49 of 118 consumer devices run it at a quantisation worth using, at an 8k context window. Another 26 can load it only by compressing the weights far enough to damage the model, marked with a warning below.
| Device | Memory | Verdict | Quantisation | Used | Speed |
|---|---|---|---|---|---|
| RTX PRO 6000 Blackwell NVIDIA | 96 GB | Runs great | Q8_0 | 43.28 GB | 194.1 t/s |
| RTX 6000 Ada Generation NVIDIA | 48 GB | Runs great | Q8_0 | 43.28 GB | 104 t/s |
| Apple M3 Ultra 96GB Apple | 96 GB | Runs great | Q8_0 | 42.98 GB | 84.4 t/s |
| Apple M3 Ultra 256GB Apple | 256 GB | Runs great | Q8_0 | 42.98 GB | 84.4 t/s |
| Apple M3 Ultra 512GB Apple | 512 GB | Runs great | Q8_0 | 42.98 GB | 84.4 t/s |
| RTX A6000 NVIDIA | 48 GB | Runs great | Q8_0 | 43.28 GB | 83.2 t/s |
| Apple M1 Ultra 64GB Apple | 64 GB | Runs great | Q8_0 | 42.98 GB | 82.4 t/s |
| Apple M1 Ultra 128GB Apple | 128 GB | Runs great | Q8_0 | 42.98 GB | 82.4 t/s |
| Apple M2 Ultra 64GB Apple | 64 GB | Runs great | Q8_0 | 42.98 GB | 82.4 t/s |
| Apple M2 Ultra 128GB Apple | 128 GB | Runs great | Q8_0 | 42.98 GB | 82.4 t/s |
| Apple M2 Ultra 192GB Apple | 192 GB | Runs great | Q8_0 | 42.98 GB | 82.4 t/s |
| Radeon PRO W7900 AMD | 48 GB | Runs great | Q8_0 | 43.28 GB | 82.2 t/s |
| Apple M4 Max 64GB Apple | 64 GB | Runs great | Q8_0 | 42.98 GB | 56.3 t/s |
| Apple M4 Max 128GB Apple | 128 GB | Runs great | Q8_0 | 42.98 GB | 56.3 t/s |
| Apple M1 Max 64GB Apple | 64 GB | Runs great | Q8_0 | 42.98 GB | 41.2 t/s |
| Apple M2 Max 64GB Apple | 64 GB | Runs great | Q8_0 | 42.98 GB | 41.2 t/s |
| Apple M2 Max 96GB Apple | 96 GB | Runs great | Q8_0 | 42.98 GB | 41.2 t/s |
| Apple M3 Max 64GB Apple | 64 GB | Runs great | Q8_0 | 42.98 GB | 41.2 t/s |
| Apple M3 Max 96GB Apple | 96 GB | Runs great | Q8_0 | 42.98 GB | 41.2 t/s |
| Apple M3 Max 128GB Apple | 128 GB | Runs great | Q8_0 | 42.98 GB | 41.2 t/s |
| NVIDIA DGX Spark 128GB NVIDIA | 128 GB | Runs great | Q8_0 | 43.28 GB | 29.6 t/s |
| Apple M4 Pro 64GB Apple | 64 GB | Runs great | Q8_0 | 42.98 GB | 28.1 t/s |
| Jetson AGX Orin 64GB NVIDIA | 64 GB | Runs well | Q8_0 | 43.28 GB | 22.2 t/s |
| Ryzen AI Max+ 395 64GB AMD | 64 GB | Runs well | Q8_0 | 43.28 GB | 22 t/s |
| Ryzen AI Max+ 395 96GB AMD | 96 GB | Runs well | Q8_0 | 43.28 GB | 22 t/s |
| Ryzen AI Max+ 395 128GB AMD | 128 GB | Runs well | Q8_0 | 43.28 GB | 22 t/s |
| Apple M4 Max 48GB Apple | 48 GB | Runs great | Q6_K | 33.53 GB | 70.2 t/s |
| Apple M3 Max 48GB Apple | 48 GB | Runs great | Q6_K | 33.53 GB | 51.4 t/s |
| Apple M4 Pro 48GB Apple | 48 GB | Runs great | Q6_K | 33.53 GB | 35.1 t/s |
| GeForce RTX 5090 NVIDIA | 32 GB | Runs great | Q5_K_M | 29.59 GB | 272.2 t/s |
| Apple M4 Max 36GB Apple | 36 GB | Just fits | Q4_K_M | 25.09 GB | 90 t/s |
| Apple M3 Max 36GB Apple | 36 GB | Just fits | Q4_K_M | 25.09 GB | 65.9 t/s |
| Apple M3 Pro 36GB Apple | 36 GB | Just fits | Q4_K_M | 25.09 GB | 24.7 t/s |
| GeForce RTX 4090 NVIDIA | 24 GB | Just fits | IQ4_XS | 22.57 GB | 192.9 t/s |
| GeForce RTX 3090 Ti NVIDIA | 24 GB | Just fits | IQ4_XS | 22.57 GB | 192.9 t/s |
| GeForce RTX 3090 NVIDIA | 24 GB | Just fits | IQ4_XS | 22.57 GB | 179.1 t/s |
| GeForce RTX 5090 Laptop NVIDIA | 24 GB | Just fits | IQ4_XS | 22.57 GB | 171.5 t/s |
| Radeon RX 7900 XTX AMD | 24 GB | Just fits | IQ4_XS | 22.57 GB | 161.3 t/s |
| RTX A5000 NVIDIA | 24 GB | Just fits | IQ4_XS | 22.57 GB | 147 t/s |
| Apple M1 Max 32GB Apple | 32 GB | Just fits | IQ4_XS | 22.27 GB | 72.8 t/s |
| Apple M2 Max 32GB Apple | 32 GB | Just fits | IQ4_XS | 22.27 GB | 72.8 t/s |
| Apple M1 Pro 32GB Apple | 32 GB | Just fits | IQ4_XS | 22.27 GB | 36.4 t/s |
| Apple M2 Pro 32GB Apple | 32 GB | Just fits | IQ4_XS | 22.27 GB | 36.4 t/s |
| Apple M5 32GB Apple | 32 GB | Just fits | IQ4_XS | 22.27 GB | 27.8 t/s |
| Apple M4 32GB Apple | 32 GB | Just fits | IQ4_XS | 22.27 GB | 21.8 t/s |
| Jetson AGX Orin 32GB NVIDIA | 32 GB | Runs great | Q3_K_M | 20.91 GB | 41.7 t/s |
| Ryzen AI Max+ 395 32GB AMD | 32 GB | Runs great | Q3_K_M | 20.91 GB | 41.4 t/s |
| Intel Core Ultra 9 288V 32GB Intel | 32 GB | Runs well | Q3_K_M | 20.91 GB | 22 t/s |
| Ryzen AI 9 HX 370 32GB AMD | 32 GB | Runs well | Q3_K_M | 20.91 GB | 20.7 t/s |
| Radeon RX 7900 XT AMD | 20 GB | Runs great | IQ3_XXS ! | 16.77 GB | 171.2 t/s |
| Apple M4 Pro 24GB Apple | 24 GB | Just fits | IQ3_XXS ! | 16.47 GB | 63.3 t/s |
| Apple M5 24GB Apple | 24 GB | Just fits | IQ3_XXS ! | 16.47 GB | 35.5 t/s |
| Apple M4 24GB Apple | 24 GB | Just fits | IQ3_XXS ! | 16.47 GB | 27.8 t/s |
| Ryzen AI 9 HX 370 24GB AMD | 24 GB | Just fits | IQ3_XXS ! | 16.77 GB | 24.7 t/s |
| Apple M2 24GB Apple | 24 GB | Just fits | IQ3_XXS ! | 16.47 GB | 23.2 t/s |
| Apple M3 24GB Apple | 24 GB | Just fits | IQ3_XXS ! | 16.47 GB | 23.2 t/s |
| GeForce RTX 5080 NVIDIA | 16 GB | Runs great | IQ2_XXS ! | 11.89 GB | 303.7 t/s |
| GeForce RTX 5070 Ti NVIDIA | 16 GB | Runs great | IQ2_XXS ! | 11.89 GB | 283.4 t/s |
| GeForce RTX 5080 Laptop NVIDIA | 16 GB | Runs great | IQ2_XXS ! | 11.89 GB | 242.9 t/s |
| GeForce RTX 4080 SUPER NVIDIA | 16 GB | Runs great | IQ2_XXS ! | 11.89 GB | 232.8 t/s |
What to buy to run Phi 3.5 MoE Instruct
The cheapest hardware that runs it at a quantisation worth using and a speed you would not resent, at an 8k context window.
Cheapest that works
GeForce RTX 3090
$700 used
IQ4_XS · uses 22.57 GB of its 24 GB · about 179.1 tokens/s
Best value
GeForce RTX 3090 Ti
$800 used
IQ4_XS · about 192.9 tokens/s · 1008 GB/s
| Device | Price | Memory | Runs it at | Speed | |
|---|---|---|---|---|---|
| GeForce RTX 3090 NVIDIA | $700 used | 24 GB | IQ4_XS | 179.1 t/s | Check price |
| GeForce RTX 3090 Ti NVIDIA | $800 used | 24 GB | IQ4_XS | 192.9 t/s | Check price |
| Apple M4 32GB Apple · whole machine | $999 | 32 GB | IQ4_XS | 21.8 t/s | Check price |
| Ryzen AI Max+ 395 64GB AMD · whole machine | $1,699 | 64 GB | Q8_0 | 22 t/s | Check price |
| Ryzen AI Max+ 395 96GB AMD · whole machine | $1,999 | 96 GB | Q8_0 | 22 t/s | Check price |
| Ryzen AI Max+ 395 128GB AMD · whole machine | $2,199 | 128 GB | Q8_0 | 22 t/s | Check price |
| GeForce RTX 5090 NVIDIA | $2,200 | 32 GB | Q5_K_M | 272.2 t/s | Check price |
| Apple M3 Ultra 256GB Apple · whole machine | $5,599 | 256 GB | Q8_0 | 84.4 t/s | Check price |
Indicative prices reviewed 2026-09-01; used prices are marketplace typical. Options that cost more than a cheaper one with no more memory and no more speed are hidden. Price links are Amazon affiliate links. Change the standard, the context window or the budget in the buying tool.
Where it sits in the catalogue
Source: microsoft/Phi-3.5-MoE-instruct on Hugging Face. Downloaded 134k times in the last month. Published 2024-08-17. Architecture figures are read from the repository's own configuration file, so they move when the model does.