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. It uses grouped-query attention with 8 KV heads for 32 query heads, which cuts the cache to 25% of what full attention would need.
About Phi 3.5 MoE Instruct
Phi 3.5 MoE Instruct is a mixture-of-experts model with 41.87B parameters, 6.64B of them active for each token, built from 32 layers with a hidden size of 4096. Its repository was published on 2024-08-17. The license listed on Hugging Face is mit. It uses grouped-query attention with 8 KV heads for 32 query heads, which cuts the cache to 25% of what full attention would need.
Memory needed, by quantization
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.
| Quantization | File size | KV cache | Total VRAM | Quality | What it costs you |
|---|---|---|---|---|---|
| Q8_0 | 41.43 GB | 1 GB | 42.71 GB | 99% | Lossless in practice. Use it when the memory is there. |
| Q6_K | 31.98 GB | 1 GB | 33.28 GB | 98% | Very close to Q8 for two thirds of the size. |
| Q5_K_M | 27.73 GB | 1 GB | 29.05 GB | 96% | The quality-first choice when Q6 will not fit. |
| Q4_K_M | 23.54 GB | 1 GB | 24.87 GB | 93% | The default. Best size-to-quality ratio for local use. |
| IQ4_XS | 20.72 GB | 1 GB | 22.06 GB | 91% | Importance-matrix 4-bit. Q4_K_S quality, smaller file. |
| Q3_K_M | 19.06 GB | 1 GB | 20.4 GB | 86% | Degradation starts to show. A way to fit one size up. |
| IQ3_XXS | 14.92 GB | 1 GB | 16.27 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. The cache grows in a straight line with every token in the window.
| Context | KV cache | Total VRAM | Fits in 8 GB | Fits in 12 GB | Fits in 24 GB |
|---|---|---|---|---|---|
| 4k | 0.5 GB | 24.37 GB | no | no | no |
| 8k | 1 GB | 24.87 GB | no | no | no |
| 16k | 2 GB | 25.87 GB | no | no | no |
| 32k | 4 GB | 27.87 GB | no | no | no |
| 64k | 8 GB | 31.87 GB | no | no | no |
| 128k | 16 GB | 39.87 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
35 of 118 consumer devices run it at a quantization worth using, at an 8k context window.
| Device | Memory | Verdict | Quantization | Used | Speed |
|---|---|---|---|---|---|
| RTX PRO 6000 Blackwell NVIDIA | 96 GB | Runs great | Q8_0 | 42.71 GB | 101.3 t/s |
| RTX 6000 Ada Generation NVIDIA | 48 GB | Runs great | Q8_0 | 42.71 GB | 54.3 t/s |
| RTX A6000 NVIDIA | 48 GB | Runs great | Q8_0 | 42.71 GB | 43.4 t/s |
| Apple M3 Ultra 96GB Apple | 96 GB | Runs great | Q8_0 | 42.81 GB | 42.8 t/s |
| Apple M3 Ultra 256GB Apple | 256 GB | Runs great | Q8_0 | 42.81 GB | 42.8 t/s |
| Apple M3 Ultra 512GB Apple | 512 GB | Runs great | Q8_0 | 42.81 GB | 42.8 t/s |
| Apple M1 Ultra 64GB Apple | 64 GB | Runs great | Q8_0 | 42.81 GB | 41.8 t/s |
| Apple M1 Ultra 128GB Apple | 128 GB | Runs great | Q8_0 | 42.81 GB | 41.8 t/s |
| Apple M2 Ultra 64GB Apple | 64 GB | Runs great | Q8_0 | 42.81 GB | 41.8 t/s |
| Apple M2 Ultra 128GB Apple | 128 GB | Runs great | Q8_0 | 42.81 GB | 41.8 t/s |
| Apple M2 Ultra 192GB Apple | 192 GB | Runs great | Q8_0 | 42.81 GB | 41.8 t/s |
| Radeon PRO W7900 AMD | 48 GB | Runs great | Q8_0 | 42.71 GB | 41.4 t/s |
| Apple M4 Max 64GB Apple | 64 GB | Runs great | Q8_0 | 42.81 GB | 28.5 t/s |
| Apple M4 Max 128GB Apple | 128 GB | Runs great | Q8_0 | 42.81 GB | 28.5 t/s |
| Apple M1 Max 64GB Apple | 64 GB | Runs well | Q8_0 | 42.81 GB | 20.9 t/s |
| Apple M2 Max 64GB Apple | 64 GB | Runs well | Q8_0 | 42.81 GB | 20.9 t/s |
| Apple M2 Max 96GB Apple | 96 GB | Runs well | Q8_0 | 42.81 GB | 20.9 t/s |
| Apple M3 Max 64GB Apple | 64 GB | Runs well | Q8_0 | 42.81 GB | 20.9 t/s |
| Apple M3 Max 96GB Apple | 96 GB | Runs well | Q8_0 | 42.81 GB | 20.9 t/s |
| Apple M3 Max 128GB Apple | 128 GB | Runs well | Q8_0 | 42.81 GB | 20.9 t/s |
| GeForce RTX 5090 NVIDIA | 32 GB | Runs great | Q5_K_M | 29.05 GB | 142 t/s |
| NVIDIA DGX Spark 128GB NVIDIA | 128 GB | Runs well | Q8_0 | 42.84 GB | 15.4 t/s |
| Apple M4 Pro 64GB Apple | 64 GB | Runs well | Q8_0 | 42.81 GB | 14.3 t/s |
| GeForce RTX 4090 NVIDIA | 24 GB | Runs great | IQ4_XS | 22.06 GB | 100.6 t/s |
| GeForce RTX 3090 Ti NVIDIA | 24 GB | Runs great | IQ4_XS | 22.06 GB | 100.6 t/s |
Show 35 more devices
| GeForce RTX 3090 NVIDIA | 24 GB | Runs great | IQ4_XS | 22.06 GB | 93.4 t/s |
| GeForce RTX 5090 Laptop NVIDIA | 24 GB | Runs great | IQ4_XS | 22.06 GB | 89.4 t/s |
| Radeon RX 7900 XTX AMD | 24 GB | Runs great | IQ4_XS | 22.06 GB | 81.3 t/s |
| RTX A5000 NVIDIA | 24 GB | Runs great | IQ4_XS | 22.06 GB | 76.7 t/s |
| Jetson AGX Orin 64GB NVIDIA | 64 GB | Runs well | Q8_0 | 42.84 GB | 11.6 t/s |
| Ryzen AI Max+ 395 64GB AMD | 64 GB | Runs well | Q8_0 | 42.84 GB | 11.2 t/s |
| Ryzen AI Max+ 395 96GB AMD | 96 GB | Runs well | Q8_0 | 42.84 GB | 11.2 t/s |
| Ryzen AI Max+ 395 128GB AMD | 128 GB | Runs well | Q8_0 | 42.84 GB | 11.2 t/s |
| Intel Core Ultra 9 288V 32GB Intel | 32 GB | Just fits | IQ4_XS | 22.12 GB | 10.5 t/s |
| Ryzen AI 9 HX 370 32GB AMD | 32 GB | Runs well | Q3_K_M | 20.46 GB | 10.5 t/s |
| Apple M4 Max 36GB Apple | 36 GB | Just fits | n/a | 29.11 GB | 40 t/s |
| Apple M1 Pro 16GB Apple | 16 GB | Just fits | n/a ! | 11.42 GB | 30.5 t/s |
| Apple M2 Pro 16GB Apple | 16 GB | Just fits | n/a ! | 11.42 GB | 30.5 t/s |
| Apple M1 Max 32GB Apple | 32 GB | Just fits | n/a | 29.11 GB | 29.3 t/s |
| Apple M2 Max 32GB Apple | 32 GB | Just fits | n/a | 29.11 GB | 29.3 t/s |
| Apple M3 Max 36GB Apple | 36 GB | Just fits | n/a | 29.11 GB | 29.3 t/s |
| Apple M4 Max 48GB Apple | 48 GB | Just fits | n/a | 42.81 GB | 28.5 t/s |
| Apple M4 Pro 24GB Apple | 24 GB | Just fits | n/a | 20.43 GB | 26.8 t/s |
| Apple M5 16GB Apple | 16 GB | Just fits | n/a ! | 11.42 GB | 23.3 t/s |
| Apple M3 Max 48GB Apple | 48 GB | Just fits | n/a | 42.81 GB | 20.9 t/s |
| Apple M4 16GB Apple | 16 GB | Just fits | n/a ! | 11.42 GB | 18.3 t/s |
| Apple M3 Pro 18GB Apple | 18 GB | Just fits | n/a ! | 16.29 GB | 17.6 t/s |
| Intel Core Ultra 9 288V 16GB Intel | 16 GB | Just fits | n/a ! | 11.45 GB | 17.3 t/s |
| Ryzen AI 9 HX 370 16GB AMD | 16 GB | Just fits | n/a ! | 11.45 GB | 16.3 t/s |
| Jetson AGX Orin 32GB NVIDIA | 32 GB | Just fits | n/a | 29.14 GB | 16.2 t/s |
| Ryzen AI Max+ 395 32GB AMD | 32 GB | Just fits | n/a | 29.14 GB | 15.7 t/s |
| Apple M2 16GB Apple | 16 GB | Just fits | n/a ! | 11.42 GB | 15.2 t/s |
| Apple M3 16GB Apple | 16 GB | Just fits | n/a ! | 11.42 GB | 15.2 t/s |
| Apple M5 24GB Apple | 24 GB | Just fits | n/a | 20.43 GB | 15 t/s |
| Apple M1 Pro 32GB Apple | 32 GB | Just fits | n/a | 29.11 GB | 14.6 t/s |
| Apple M2 Pro 32GB Apple | 32 GB | Just fits | n/a | 29.11 GB | 14.6 t/s |
| Apple M4 Pro 48GB Apple | 48 GB | Just fits | n/a | 42.81 GB | 14.3 t/s |
| Apple M4 24GB Apple | 24 GB | Just fits | n/a | 20.43 GB | 11.8 t/s |
| Apple M5 32GB Apple | 32 GB | Just fits | n/a | 29.11 GB | 11.2 t/s |
| Apple M3 Pro 36GB Apple | 36 GB | Just fits | n/a | 29.11 GB | 11 t/s |
How it compares
Other models in the same family or of a similar size, on a 12 GB card and a 24 GB card, each at the best quantization that fits.
| Model | Weights at 4 bits | GeForce RTX 3060 12GB | GeForce RTX 4090 |
|---|---|---|---|
| Phi 3.5 MoE Instruct | 23.5 GB | 13.1 t/s at Q4_K_M | 100.6 t/s at IQ4_XS |
| Phi 4 | 8.2 GB | 27.8 t/s at Q4_K_M | 47.7 t/s at Q8_0 |
| Phi 4 Mini Instruct | 2.2 GB | 55.2 t/s at Q8_0 | 154.5 t/s at Q8_0 |
| Phi 3 Mini 4k Instruct | 2.1 GB | 59.8 t/s at Q8_0 | 167.3 t/s at Q8_0 |
| Gemma 4 31B Instruct | 17.6 GB | does not fit well | 35.5 t/s at Q4_K_M |
| Qwen3 32B | 18.4 GB | does not fit well | 37.1 t/s at Q4_K_M |
What to buy to run Phi 3.5 MoE Instruct
The cheapest hardware that runs it at a quantization worth using and a speed you would not resent, at an 8k context window.
Cheapest that works
GeForce RTX 2060 12GB
$160 used
IQ4_XS · uses 22.06 GB of its 12 GB · about 15.3 tokens/s
Best value
GeForce RTX 3090
$700 used
IQ4_XS · about 93.4 tokens/s · 936 GB/s
If it just has to run
GeForce GTX 1650
$100 used
Q4_K_M · about 7.9 tokens/s. Cheaper, and slower to answer.
| Device | Price | Memory | Runs it at | Speed | |
|---|---|---|---|---|---|
| GeForce GTX 1080 Ti NVIDIA | $150 used | 11 GB | IQ4_XS | 15.9 t/s | Check price |
| GeForce RTX 2060 12GB NVIDIA | $160 used | 12 GB | IQ4_XS | 15.3 t/s | Check price |
| GeForce RTX 2080 Ti NVIDIA | $250 used | 11 GB | IQ4_XS | 16.7 t/s | Check price |
| Arc A770 16GB Intel | $280 used | 16 GB | Q4_K_M | 16.9 t/s | Check price |
| GeForce RTX 5060 Ti 16GB NVIDIA | $430 | 16 GB | Q4_K_M | 17.1 t/s | Check price |
| GeForce RTX 4060 Ti 16GB NVIDIA | $450 | 16 GB | IQ4_XS | 17.9 t/s | Check price |
| Radeon RX 7800 XT AMD | $480 | 16 GB | Q4_K_M | 18 t/s | Check price |
| GeForce RTX 5070 NVIDIA | $560 | 12 GB | IQ4_XS | 18.1 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 catalog
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.
Questions people ask about Phi 3.5 MoE Instruct
How much VRAM does Phi 3.5 MoE Instruct need?
At Q4_K_M with an 8k context window, Phi 3.5 MoE Instruct needs about 24.87 GB: 23.47 GB of weights, 1 GB of KV cache and 0.41 GB of overhead. At the highest-quality file it needs about 42.71 GB.
What is the cheapest GPU that runs Phi 3.5 MoE Instruct?
The smallest consumer device in our catalog that fits it is the GeForce RTX 4090 with 24 GB, running it at IQ4_XS at roughly 100.6 tokens per second.
How much does a longer context window add for Phi 3.5 MoE Instruct?
On the memory side, a 32k window adds 4.0 GB of KV cache on top of the weights, and the cache grows in a straight line with every token. The longest window the model supports natively is 128k tokens.