All models › microsoft
Phi-tiny-MoE-instruct
microsoft · open license · on Arena since 23 Jun 2025
Download No Arena votes
Arena score—Not voted on Arena yet
Size3.8 B1.3 B active per token (mixture of experts)
4-bit download2.3 GBweights only, Q4_K_M
Max context4Ktokens, from the model config
Where to use it for free
- Download and run it at home: official weights · GGUF files for Ollama and LM Studio
No Arena votes yet
Arena has not voted on this model. It is listed because it is a popular small model on its lab's official Hugging Face page (38K downloads a month).
Speed on popular hardware
| Hardware | GPU memory | Fits? | Speed |
|---|---|---|---|
| RTX 3060 12 GB | 12 GB | Fits in GPU | 86 tok/s |
| RTX 4060 Ti 16 GB | 16 GB | Fits in GPU | 69 tok/s |
| RTX 3090 | 24 GB | Fits in GPU | 223 tok/s |
| RTX 4090 | 24 GB | Fits in GPU | 240 tok/s |
| RTX 5090 | 32 GB | Fits in GPU | 427 tok/s |
| RX 7900 XTX | 24 GB | Fits in GPU | 229 tok/s |
| 2 × RTX 3090 | 48 GB | Fits in GPU | 217 tok/s |
| 4 × RTX 3090 | 96 GB | Fits in GPU | 205 tok/s |
| Mac M4 Max 128 GB | 96 GB | Fits in GPU | 130 tok/s |
| Ryzen AI Max+ 395 128 GB | 96 GB | Fits in GPU | 61 tok/s |
| 8 × H100 SXM | 640 GB | Fits in GPU | 660 tok/s |
Estimate at 4-bit with 32K context, 64 GB of DDR5 for offload. Speed ≈ memory bandwidth × 70% ÷ data read per token; real numbers vary by about ±30%.
GPU memory needed
| Level | Quality vs full |
|---|---|
| Q3_K_M | +10.5% noticeable loss |
| Q4_K_M | +2.8% small loss |
| Q5_K_M | +0.9% very small loss |
| Q6_K | +0.3% near-lossless |
| Q8_0 | ≈ 0% same as full |
| F16 | ≈ 0% same as full |
Weights + context cache (from the model's architecture) + runtime. Quality: increase in perplexity against the full model, from llama.cpp tests, scaled by model size. Simulate your own setup.