Phi 3.5 MoE Instruct — Hardware Requirements & GPU Compatibility
ChatCode
Phi-3.5-MoE-instruct is Microsoft's Mixture-of-Experts model in the Phi-3.5 line, combining 16 experts of 3.8 billion parameters each, about 42 billion total, with only around 6.6 billion active per token. It is an instruction-tuned model refined with supervised fine-tuning, PPO, and DPO, built for multilingual reasoning tasks including code, math, and logic in memory- or latency-constrained deployments. It supports a 128K token context window and ships under the MIT license, allowing free commercial and research use. Because only about 6.6 billion parameters activate per token, it runs noticeably faster than its 42-billion-parameter size implies, though the full model still needs roughly 24 GB of memory at 4-bit quantization since all expert weights stay resident.
Usable on a 24 GB GPU
Runs from 16 GB (IQ2_S · 16K)Best on 24 GB (Q3_K_L · ~141 tok/s · 16K)Macs from 18 GB (IQ2_XXS · ~41 tok/s)
MIT
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Phi 3.5 MoE Instruct
One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.
Recommended
ollama run hf.co/bartowski/Phi-3.5-MoE-instruct-GGUF:Q4_K_MQ4_K_M · needs ~27.6 GB · 16K context
Download: Phi-3.5-MoE-instruct-Q4_K_M.gguf25.35 GBbartowski
Ways to runOther ways
each opens step-by-stepRun in cloud
Runs on A40 (48 GB) (20 GB headroom) · Q4_K_M
- Generation speed
- ~93 tok/s
- generation speed
- Cost per 1M output tokens
- $1.05
- per 1M output tokens
How Much VRAM Does Phi 3.5 MoE Instruct Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 18.4 GB | 22.4 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 32K context · ~161 tok/s | |
| 18.9 GB | 22.9 GB | 3-bit small — 3-bit small quantization | Fits: 32K context · ~158 tok/s | |
| 21.0 GB | 25 GB | 3-bit medium — 3-bit medium quantization | Fits: 16K context · ~146 tok/s | |
| 21.5 GB | 25.5 GB | 4-bit · legacy — 4-bit legacy quantization | Fits: 16K context · ~144 tok/s | |
| 25.7 GB | 29.7 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Does not fit: needs 27.6 GB | |
| 30.4 GB | 34.4 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Does not fit: needs 32.3 GB | |
| 35.1 GB | 39.1 GB | 6-bit, very good — 6-bit quantization, very good quality | Does not fit: needs 37 GB | |
| 42.4 GB | 46.5 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Does not fit: needs 44.3 GB |
Selected Q4_K_M
4-bit medium quantization27.6GBat 16K context
- Weights
- 25.1 GB
- KV cache
- 2.2 GB
- Overhead
- 0.3 GB
Doesn't fit RTX 4090 (reference) · needs 3.6 GB more, Q3_K_L fits
Or .
Which Hardware Can Run Phi 3.5 MoE Instruct?
Q4_K_M · 25.7 GBPhi 3.5 MoE Instruct (Q4_K_M) requires 25.7 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 34+ GB is recommended. Using the full 131K context window can add up to 16.9 GB, bringing total usage to 42.6 GB. 1 GPU can run it, including NVIDIA GeForce RTX 5090. 98 devices with unified memory can run Phi 3.5 MoE Instruct, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Studio M5 Max (36 GB).
20 GB
Q4_K_M · best fit IQ3_XS (19.7 GB)
1 GPU
RX 7900 XT~81Simulate output on RX 7900 XT32–40 GB
fits · 4.4 GB free · up to 32K ctx
1 GPU · 3 Mac classes · 2 device families
RTX 5090~190Simulate output on RTX 5090M5 Max36 GB~53Simulate output on Mac Studio M5 Max (36 GB)M532 GB~23Simulate output on MacBook Air 13" M5 (32 GB)Orin~31Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~20Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at Q4_K_M
How good is Phi 3.5 MoE Instruct?
llmrun Score blends public benchmark boards into one number; 100 = the average open model.
Not enough data yet for an llmrun Score.
A Score needs results on enough independent boards. How the Composite Score is built
Benchmarks
Here are all 3 boards that have results for this model, sorted by their weight in the llmrun Score; badges show who ran the test.
Alternatives to Phi 3.5 MoE Instruct
Phi 3 family · 15 models →Similar models
2| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| FlintRecursiveSI | — | 23.5 GB | ~205 | same size |
| Falcon 40B InstructTII UAE | — | 23.8 GB | ~33 | same size |
Specifications
- Publisher
- Microsoft
- Family
- Phi 3
- Parameters
- 41.9B
- Active parameters
- 6.6B per token
- Architecture
- PhiMoEForCausalLM
- Context Length
- 131,072 tokens
- Vocabulary Size
- 32,064
- Release Date
- 2024-08-17
- License
- MIT
Frequently Asked Questions
How much VRAM does Phi 3.5 MoE Instruct need?
Phi 3.5 MoE Instruct requires 25.7 GB of VRAM at Q4_K_M, or 84.3 GB at BF16. Full 131K context adds up to 16.9 GB (42.6 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 41.9B × 4.8 bits ÷ 8 = 25.1 GB
KV Cache + Overhead ≈ 0.6 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 17.5 GB (at full 131K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 4090 run Phi 3.5 MoE Instruct?
Yes, at IQ4_XS (23.1 GB) or lower. Higher quantizations like Q4_1 (24.1 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for Phi 3.5 MoE Instruct?
For Phi 3.5 MoE Instruct, Q4_K_M (25.7 GB) offers the best balance of quality and VRAM usage. Q4_K_L (26.2 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 12.1 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Phi 3.5 MoE Instruct on a Mac?
Yes — MacBook Pro 14" M3 Pro (18 GB) and 108 other Macs can run Phi 3.5 MoE Instruct, the smallest at IQ2_XXS. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 25.7 GB of usable unified memory (RAM minus macOS overhead), from Mac Studio M1 Max (32 GB).
Can I run Phi 3.5 MoE Instruct locally?
Yes — Phi 3.5 MoE Instruct can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Phi 3.5 MoE Instruct?
At Q4_K_M, Phi 3.5 MoE Instruct can reach ~138 tok/s on AMD Instinct MI350X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)
Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.
Example: NVIDIA B200 → 4.1 GB active ÷ (8000 × 0.65) = 0.784 ms, plus 32 layers × 0.055 ms = 1.760 ms, so 1000 ÷ 2.544 ms = ~393 tok/s
Estimated speed at Q4_K_M (25.7 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Phi 3.5 MoE Instruct?
At Q4_K_M, the download is about 25.12 GB. The full-precision BF16 version is 83.75 GB. The smallest option (IQ2_XXS) is 11.52 GB.
Which GPUs can run Phi 3.5 MoE Instruct?
1 consumer GPU can run Phi 3.5 MoE Instruct at Q4_K_M. Top options include NVIDIA GeForce RTX 5090.
Which devices can run Phi 3.5 MoE Instruct?
98 devices with unified memory can run Phi 3.5 MoE Instruct at Q4_K_M (25.7 GB), including ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB), GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.