Phi-3 Mini 3.8B Instruct
Phi · MIT
Punches above its weight: a 3.8B model that rivals many 7B models on reasoning benchmarks. MIT license. Well suited for resource-constrained setups or as a fast secondary model.
- Parameters
- 3.82B
- Architecture
- Dense
- Context
- 128,000 tokens
- Released
- 2024-04-23
- Engines
- llama.cpp, ollama, ONNX Runtime
- Builder Tools
- Continue, LM Studio
Parameters
3.82B
VRAM
3 GB
Context
125K
Formats
3
GPUs
20
Phi-3 Mini 3.8B Instruct (3.82B) requires 3 GB VRAM at recommended quality (Q5_K_M). At efficient quality (Q4_K_M), it fits in 2.6 GB VRAM, making it compatible with the NVIDIA RTX 4060 Laptop (40-60W). On NVIDIA Grace Blackwell Ultra GB300, expect approximately 550 tok/s at Q8_0. For the best experience, Starter AI Desktop ($582) is recommended.
Source: OwnRig methodology
3 GB
Q5_K_M
2.3 GB
125K tokens
Chat
VRAM Requirements
| Quality | Quantization | VRAM | File Size |
|---|---|---|---|
| full | Q8_0 | 4.5 GB | 3.8 GB |
| recommended | Q5_K_M | 3 GB | 2.3 GB |
| efficient | Q4_K_M | 2.6 GB | 1.9 GB |
Context Length Impact
KV cache VRAM at Q5_K_M quality. Longer context = more memory.
| Context | KV Cache | Total VRAM |
|---|---|---|
| 2K | 102 MB | 3.1 GB |
| 4K | 102 MB | 3.1 GB |
| 8K | 307 MB | 3.3 GB |
| 16K | 512 MB | 3.5 GB |
| 32K | 1 GB | 4 GB |
| 64K | 2 GB | 5 GB |
Compatible GPUs
20 devices| NVIDIA Grace Blackwell Ultra GB300 | Q8_0 | 550 tok/s | Excellent |
| NVIDIA GeForce RTX 3060 12GB | Q8_0 | 60 tok/s | Excellent |
| NVIDIA GeForce RTX 3080 10GB | Q8_0 | 130 tok/s | Excellent |
| NVIDIA GeForce RTX 4060 8GB | Q5_K_M | 52 tok/s | Excellent |
| NVIDIA GeForce RTX 4070 Super | Q8_0 | 95 tok/s | Excellent |
| NVIDIA GeForce RTX 4070 Ti 12GB | Q8_0 | 78 tok/s | Excellent |
| NVIDIA RTX 4080 Laptop (120-150W) | Q8_0 | 55 tok/s | Excellent |
| NVIDIA GeForce RTX 4090 | Q8_0 | 130 tok/s | Excellent |
| AMD Radeon Pro W7900 | Q8_0 | 140 tok/s | Excellent |
| NVIDIA RTX PRO 6000 Blackwell | Q8_0 | 257 tok/s | Excellent |
| NVIDIA RTX PRO 6000 Blackwell Max-Q | Q8_0 | 236 tok/s | Excellent |
| NVIDIA GeForce RTX 5060 8GB | Q5_K_M | 60 tok/s | Excellent |
| Apple M3 Pro (18GB Unified) | Q8_0 | 32 tok/s | Good |
| NVIDIA RTX 4060 Laptop (40-60W) | Q5_K_M | 31 tok/s | Good |
| NVIDIA RTX 4070 Laptop (80-115W) | Q5_K_M | 36 tok/s | Good |
| AMD Radeon RX 7600 | Q5_K_M | 41 tok/s | Good |
| AMD Radeon RX 7900 XTX | Q8_0 | 112 tok/s | Good |
| AMD Radeon RX 9070 | Q8_0 | – | Acceptable |
| AMD Radeon RX 9060 XT 16GB | Q8_0 | – | Acceptable |
| AMD Radeon RX 9060 XT 8GB | Q8_0 | – | Acceptable |
Showing 20 of 20 entries
Builder Context
Phi-3 Mini 3.8B Instruct is commonly used with Continue, LM Studio. For an AI coding workflow, pair it with an embedding model like nomic-embed-text for local RAG.
Recommended Builds
Complete PC builds that can run Phi-3 Mini 3.8B Instruct.
Frequently Asked Questions
- How much VRAM does Phi-3 Mini 3.8B Instruct need?
- Phi-3 Mini 3.8B Instruct requires 3 GB VRAM at recommended quality (Q5_K_M). At lower quality settings, it can fit in as little as 2.6 GB.
- What is the best GPU for Phi-3 Mini 3.8B Instruct?
- The NVIDIA Grace Blackwell Ultra GB300 delivers the best performance for Phi-3 Mini 3.8B Instruct, achieving 550 tok/s at Q8_0 with an excellent rating.
- What quantization should I use for Phi-3 Mini 3.8B Instruct?
- For the best quality, use Q5_K_M (3 GB VRAM). If your GPU has limited VRAM, Q4_K_M (2.6 GB) is the most efficient option with acceptable quality.
- Is Phi-3 Mini 3.8B Instruct good for coding?
- Yes. Phi-3 Mini 3.8B Instruct is used with Continue, LM Studio for local AI coding. For the best coding experience, pair it with an embedding model for local RAG.
Data confidence: verified. Source
VRAM requirements are calculated from model parameters and may vary by inference engine, context length, and batch size. Performance estimates are based on community benchmarks and should be verified for your specific configuration.Phi is a trademark of its respective owner. OwnRig is not affiliated with or endorsed by the model creator.