Phi-3 Models by Microsoft Research
Details
Capabilities
About
The Phi-3 family, developed by Microsoft, consists of small language models (SLMs) optimized for Azure AI 1. These models are known for their capability and cost-effectiveness, outperforming larger models in tasks such as language processing, reasoning, coding, and math 1. The Phi-3 lineup includes models like the Phi-3-mini with 3.8 billion parameters, and the Phi-3-small and Phi-3-medium, each with 7 billion and 14 billion parameters, respectively 1. Notably, the Phi-3-mini supports up to a 128K token context window with minimal quality impact, a feature rare for models of its size 1. These models are instruction-tuned for straightforward usage and are optimized for various processing platforms, including GPUs, CPUs, and even mobile hardware 1. Despite their prowess, the Phi-3 models might perform slightly below larger models on factual knowledge tests due to their relatively smaller size 1.
Current Variants
Use-when guidance is based on each model's tracked capabilities, context window, release date, and replacement status.
Use when the workload needs 128k context and 3.8B parameters.
Use when the workload needs 128k context, 4.1B parameters, and multimodal inputs.
Use when the workload needs 128k context and 14B parameters.
Use when the workload needs 4k context, 14B parameters, and structured outputs.
Use when the workload needs 128k context and 7B parameters.
Use when the workload needs 8k context and 7B parameters.
Use when the workload needs 128k context, 4.2B parameters, and multimodal inputs.
Use when the workload needs 4k context, 3.8B parameters, and structured outputs.
Use when the workload needs 128k context, 7B parameters, and structured outputs.
Use when the workload needs 128k context and 3.8B parameters.
Use when the workload needs 4k context and 3.8B parameters.
Use when the workload needs 4k context and 3.8B parameters.
Use when the workload needs 4k context and 14B parameters.
| Model | Use when | Released | Signals | Status |
|---|---|---|---|---|
| Phi 3.5 Mini Instruct | Use when the workload needs 128k context and 3.8B parameters. | 2024-08 | 128k context3.8B parameters | Current |
| Phi 3.5 MoE Instruct | Use when the workload needs 128k context. | 2024-08 | 128k context | Current |
| Phi 3.5 Vision Instruct | Use when the workload needs 128k context, 4.1B parameters, and multimodal inputs. | 2024-08 | 128k context4.1B parametersmultimodal inputs | Current |
| Phi-3 Silica | Use when the workload needs 3.3B parameters. | 2024-06 | 3.3B parameters | Current |
| Phi-3 Medium 128K | Use when the workload needs 128k context and 14B parameters. | 2024-05 | 128k context14B parameters | Current |
| Phi-3 Medium 4K | Use when the workload needs 4k context, 14B parameters, and structured outputs. | 2024-05 | 4k context14B parametersstructured outputs | Current |
| Phi-3 Small 128K | Use when the workload needs 128k context and 7B parameters. | 2024-05 | 128k context7B parameters | Current |
| Phi-3 Small 8K | Use when the workload needs 8k context and 7B parameters. | 2024-05 | 8k context7B parameters | Current |
| Phi-3 Vision | Use when the workload needs 128k context, 4.2B parameters, and multimodal inputs. | 2024-05 | 128k context4.2B parametersmultimodal inputs | Current |
| DeepInfra Phi 3 Mini 4K Instruct | Use when the workload needs 4k context, 3.8B parameters, and structured outputs. | 2024-05 | 4k context3.8B parametersstructured outputs | Current |
| DeepInfra Phi 3 Small 128K Instruct | Use when the workload needs 128k context, 7B parameters, and structured outputs. | 2024-05 | 128k context7B parametersstructured outputs | Current |
| Phi-3 Mini 128K | Use when the workload needs 128k context and 3.8B parameters. | 2024-04 | 128k context3.8B parameters | Current |
| Phi-3 Mini 4k | Use when the workload needs 4k context and 3.8B parameters. | 2024-04 | 4k context3.8B parameters | Current |
| Phi-3 Mini | Use when the workload needs 4k context and 3.8B parameters. | 2024-04 | 4k context3.8B parameters | Current |
| Phi-3 Medium | Use when the workload needs 4k context and 14B parameters. | 2024-04 | 4k context14B parameters | Current |
Release Timeline
4 release groupsSpecifications(15 models)
| Model | Released | Context | Parameters | Vision | Multimodal | Structured Outputs |
|---|---|---|---|---|---|---|
| Phi 3.5 Mini Instruct | 2024-08 | 128k | 3.8B | No | No | No |
| Phi 3.5 MoE Instruct | 2024-08 | 128k | 16x3.8B (42B, 6.6B active) | No | No | No |
| Phi 3.5 Vision Instruct | 2024-08 | 128k | 4.1B | Yes | Yes | No |
| Phi-3 Silica | 2024-06 | — | 3.3B | No | No | No |
| Phi-3 Medium 128K | 2024-05 | 128k | 14B | No | No | No |
| Phi-3 Medium 4K | 2024-05 | 4k | 14B | No | No | Yes |
| Phi-3 Small 128K | 2024-05 | 128k | 7B | No | No | No |
| Phi-3 Small 8K | 2024-05 | 8k | 7B | No | No | No |
| Phi-3 Vision | 2024-05 | 128k | 4.2B | Yes | No | No |
| DeepInfra Phi 3 Mini 4K Instruct | 2024-05 | 4k | 3.8B | No | No | Yes |
| DeepInfra Phi 3 Small 128K Instruct | 2024-05 | 128k | 7B | No | No | Yes |
| Phi-3 Mini 128K | 2024-04 | 128k | 3.8B | No | No | No |
| Phi-3 Mini 4k | 2024-04 | 4k | 3.8B | No | No | No |
| Phi-3 Mini | 2024-04 | 4k | 3.8B | No | No | No |
| Phi-3 Medium | 2024-04 | 4k | 14B | No | No | No |
Available From(6 providers)
Pricing
| Model | Provider | Input / 1M | Output / 1M | Type |
|---|---|---|---|---|
| DeepInfra Phi 3 Mini 4K Instruct | DeepInfra | $0.05 | $0.15 | Serverless |
| Phi-3 Mini 128K | Replicate API | $0.05 | $0.25 | Serverless |
| Phi-3 Mini 4k | Replicate API | $0.05 | $0.25 | Serverless |
| Phi-3 Mini 128K | Fireworks AI | $0.1 | $0.1 | Provisioned |
| Phi-3 Medium 4K | DeepInfra | $0.14 | $0.41 | Serverless |
| Phi-3 Vision | Fireworks AI | $0.2 | $0.2 | Serverless |
| Phi-3 Mini 4k | Microsoft Foundry | $0.28 | $0.84 | Serverless |
| Phi-3 Vision | Microsoft Foundry | $0.28 | $0.84 | Provisioned |
| Phi-3 Mini 128K | Microsoft Foundry | $0.3 | $0.9 | Serverless |
| Phi-3 Small 8K | Microsoft Foundry | $0.32 | $0.96 | Serverless |
| Phi-3 Small 128K | Microsoft Foundry | $0.35 | $1.05 | Serverless |
| Phi-3 Medium 4K | Microsoft Foundry | $0.45 | $1.35 | Serverless |
| DeepInfra Phi 3 Small 128K Instruct | DeepInfra | $0.45 | $0.65 | Serverless |
| Phi-3 Medium 128K | Microsoft Foundry | $0.5 | $1.5 | Serverless |
| Phi 3.5 MoE Instruct | Fireworks AI | $0.5 | $0.5 | Serverless |
| Phi 3.5 Mini Instruct | Fireworks AI | $0.9 | $0.9 | Serverless |
Popular comparisons in this family
- Llama 3.2 1B Instruct vs Phi-3 Mini 4k348
- Mistral Large 2 vs Phi-3 Mini 4k158
- Llama 3 8B Instruct vs Phi-3 Mini 4k146
- Llama 3 70B Instruct vs Phi-3 Mini 4k113
- Llama 3 8B Instruct vs Phi 3.5 Mini Instruct73
- o3 Mini vs Phi-3 Mini 4k62
- Phi 3.5 Vision Instruct vs Phi-4 Mini Flash Reasoning53
- GPT-5.5 vs Phi-3 Mini 4k49
- Mixtral 8x7B vs Phi-3 Mini 4k45
- DeepSeek V4 Pro vs Phi-3 Mini 4k45
Frequently Asked Questions
- What is Phi-3 used for?
- Phi-3 is used for vision and multimodal work, structured outputs, and coding. The family description and listed model capabilities point to those workloads as the best fit.
- How does Phi-3 compare to Harrier?
- Phi-3 by Microsoft Research is strongest where you need vision and multimodal work, while Harrier by Microsoft Research is the closest related family to check for embedding. Phi-3 has 15 listed variants and reaches up to 128k context, while Harrier reaches up to 33k context, so compare the specs and pricing tables before choosing a production model.
- Which Phi-3 model should I use?
- For the lowest listed input price, start with DeepInfra Phi 3 Mini 4K Instruct through DeepInfra at $0.05/1M input tokens. For the most capable/latest local choice, evaluate Phi 3.5 Vision Instruct with 128k context and multimodal inputs.
Related Model Families
Models(15)
Phi 3.5 Mini Instruct
Phi 3.5 MoE Instruct
Phi 3.5 Vision Instruct
Phi-3 Silica
Phi-3 Medium 128K
Phi-3 Medium 4K
Phi-3 Small 128K
Phi-3 Small 8K
Phi-3 Vision
DeepInfra Phi 3 Mini 4K Instruct
DeepInfra Phi 3 Small 128K Instruct
Phi-3 Mini 128K
Phi-3 Mini 4k
Phi-3 Mini
Phi-3 Medium