Phi3-3.8|7|14B
FreeCompact 3.8B instruct model from Microsoft, MIT licensed, strong on reasoning benchmarks.
FreeFree tier
Inputs: textOutputs: text
About Phi3-3.8|7|14B
Phi-3-mini-4k-instruct is a compact, instruction-tuned language model developed by Microsoft, featuring 3.8 billion parameters and a 4096-token context window. Released under the permissive MIT license, it is designed for text generation, conversational AI, code generation, and reasoning tasks. The model achieves competitive benchmark scores, including 85.7% on GSM8K (8-shot CoT) and 45.66 on MMLU-Pro, while supporting English and French. It is built on the Transformers library and is available for free download and inference on Hugging Face, with over 30 million all-time downloads and strong community engagement.
Key Features
3.8 billion parameters (BF16) for efficient deployment
4096-token context window (4K)
Instruction-tuned for conversational and task-oriented text generation
Achieves 85.7% on GSM8K (8-shot CoT) and 45.66 on MMLU-Pro
Supports English and French languages
Released under permissive MIT license
Built on Hugging Face Transformers with safetensors format
Optimized for text-generation-inference and inference endpoints
Multilingual tags include English (en) and French (fr)
Over 30 million all-time downloads and 1,445 likes
Pros & Cons
Pros
- Strong performance on math and reasoning benchmarks for its size
- Permissive MIT license allows commercial use and modification
- Compact 3.8B parameter size enables efficient inference on consumer hardware
- Active community and extensive documentation on Hugging Face
- Supports both English and French out-of-the-box
- Sharded safetensors for optimized loading and memory usage
Cons
- 4K context window may be insufficient for long-document or very long conversation tasks
- Primarily focused on English and French; limited multilingual coverage
- Not as powerful as larger models (e.g., 7B+, 13B) for complex reasoning or creative tasks
- Requires familiarity with Hugging Face Transformers and text-generation pipeline
Best For
Conversational AI and chatbot developmentCode generation and coding assistanceMathematical reasoning and problem solving (GSM8K, MMLU-Pro)Text generation for creative writing, summarization, and Q&AEducational tutoring and knowledge retrievalMulti-turn dialogue systemsFine-tuning for domain-specific NLP tasks
FAQ
What is the parameter size of Phi-3-mini-4k-instruct?
It has approximately 3.8 billion parameters (BF16).
What license is it released under?
It is released under the MIT license, which is very permissive for commercial and non-commercial use.
What are the benchmark results?
Phi-3-mini-4k-instruct achieves 85.7% on GSM8K (8-shot CoT) and 45.66 on MMLU-Pro, as reported on its Hugging Face page.
What context length does it support?
The model supports a context window of 4096 tokens (4K).
What languages does it support?
The model is tagged for English (en) and French (fr), indicating support for these languages.
Where can I download or try the model?
It is available on Hugging Face at https://huggingface.co/microsoft/Phi-3-mini-4k-instruct.