Microsoft Phi models, generation by generation
Microsoft's small-model line from Phi-1 to Phi-4 and the Phi MoE models: sizes, dense vs MoE, context lengths, MIT license, and which to use today.
What the line is
Phi is Microsoft's family of small language models, trained heavily on synthetic and filtered data and published on Hugging Face under the microsoft organization. The point of the line is capability per parameter: the models covered here range from 1.3B to 14B dense, plus small MoE models. Every model card checked for this page lists the MIT license, so there is no custom license to read for derivatives or commercial use.
The line starts with small base models for code and reasoning, moves to instruct-tuned chat models in the Phi-3 generation, and in Phi-4 splits into a 14B model, a 3.8B mini model, reasoning variants and a multimodal model. A separate line of small mixture-of-experts models sits beside it.
Generations in order
Phi-1 (September 2023)
Phi-1 is a 1.3B-parameter Transformer specialized for basic Python coding, trained on 54B tokens (7B unique), per its card. Phi-1.5, a 1.3B model trained on the same data plus synthetic NLP text for common sense and reasoning, lives in the same hub. The Hugging Face repositories for both were created on September 10, 2023.
Phi-2 (December 2023)
Phi-2 is a 2.7B Transformer with a 2,048-token context, trained on 1.4T tokens. The repository was created on December 13, 2023.
Phi-3 (April to May 2024)
Phi-3 is the first generation in the line with instruct-tuned releases in several sizes, all dense decoder-only Transformers:
- Mini: 3.8B parameters, in 4K and 128K context variants (repositories created April 22, 2024).
- Small: 7B, with 8K and 128K variants, using alternating dense and blocksparse attention (created May 7, 2024).
- Medium: 14B, with 4K and 128K variants (created May 7, 2024).
The cards state SFT plus DPO alignment. Phi-3 also includes a vision model, which the hub covers.
Phi-3.5 (August 2024)
Phi-3.5 refreshed the family in August 2024. Phi-3.5-mini is a 3.8B dense model with a 128K context window. Phi-3.5-MoE is a mixture-of-experts model: 16 experts of 3.8B each, 6.6B active parameters with two experts selected, and a 128K window. The card is candid that the MoE model is limited by its size on tasks needing a lot of factual knowledge and suggests pairing it with retrieval.
Phi-4 (December 2024 onward)
Phi-4 starts with the 14B dense model. Its card gives a December 12, 2024 release date, a 16K context, and 9.8T training tokens. Variants followed:
- Phi-4-mini-instruct (repository created February 19, 2025): 3.8B dense, 128K context, with a 200K vocabulary, grouped-query attention and shared input and output embeddings compared with Phi-3.5-mini.
- Phi-4-multimodal-instruct (February 24, 2025): a multimodal model with a 128K context window.
- Phi-4-reasoning: 14B, built on the same base as Phi-4, with a 32K context and a stated release date of April 30, 2025. The card says inference should use temperature 0.8, top-k 50 and top-p 0.95, with up to 32,768 new tokens for hard problems. Phi-4-reasoning-plus and Phi-4-mini-reasoning are in the same hub.
Side line: Phi MoE
Phi MoE holds the SlimMoE models, published in June 2025 and distilled from the base model shared by Phi-3.5-MoE and GRIN-MoE:
- Phi-mini-MoE-instruct: 7.6B total, 2.4B active.
- Phi-tiny-MoE-instruct: 3.8B total, 1.1B active.
Both cards list a 4K context and 400B training tokens. The context is short, so these suit short prompts, not documents.
Which generation to use today
The picks below come from sizes, context and license, not benchmark claims.
- General chat or assistant on one GPU: Phi-4 at 14B is the largest dense model covered here, but its 16K context is the limit. Phi-4-mini-instruct at 3.8B keeps a 128K window.
- Long documents: Phi-4-mini-instruct, Phi-3.5-mini or the 128K variants of Phi-3. Avoid Phi-2, the Phi-4 14B model and the MoE models above for this.
- Step-by-step reasoning: Phi-4-reasoning (32K context), with the sampling settings from its card.
- Non-text input: Phi-4-multimodal-instruct, with a 128K window per its card.
- Tiny or edge-scale experiments: Phi-1, Phi-1.5 and Phi-2 are small but old and have short contexts; the MoE models above are a newer small option with 4K.
- Licensing: all of the above are MIT-licensed according to their cards.
You can rent GPUs to run any of these on Aquanode GPUs.
Phi generations
Every Phi generation we track, oldest first. Each hub lists all of its models with VRAM at native, FP8 and INT4 precision.
| Hub | Models | Sizes | Smallest native VRAM | Cheapest live fit for it | Est. $/hr |
|---|---|---|---|---|---|
| Phi-1 | 2 | 1.4B | 3.2 GB | RTX 3070 | $0.088/hr |
| Phi-2 | 1 | 2.8B | 6.2 GB | RTX 3070 | $0.088/hr |
| Phi-3 | 6 | 3.8B to 14.0B | 8.5 GB | RTX 3060 | $0.110/hr |
| Phi-3.5 | 4 | 3.8B to 41.9B | 8.5 GB | RTX 3060 | $0.110/hr |
| Phi-4 | 7 | 3.8B to 14.7B | 8.6 GB | RTX 3060 | $0.110/hr |
Smallest native VRAM is the lowest requirement among the publisher's own checkpoints in the hub, at the precision they are published in; see the methodology.
Other Phi lines
Side lines from Microsoft: specialised or companion models outside the main Phi generations.
Run and fine-tune Phi
Per-generation guides: engine support, chat template and context flags, then fine-tune memory for LoRA and QLoRA.
- Phi-4: how to run, fine-tune
Phi compared with other models
Side-by-side VRAM, context length and license for models of a similar size.
- phi-4 vs Qwen2.5-14B-Instruct
- Phi-3-mini-4k-instruct vs Qwen2.5-3B-Instruct
- Phi-3.5-MoE-instruct vs Qwen-72B
- phi-2 vs Qwen3-0.6B
- phi-4 vs Qwen2.5-Coder-14B-Instruct
- Phi-3-mini-4k-instruct vs Qwen3-4B
- Llama-3.3-70B-Instruct vs Phi-3.5-MoE-instruct
- phi-2 vs Qwen2.5-1.5B-Instruct
- phi-4 vs Qwen3-14B
- Phi-3-mini-4k-instruct vs Qwen3-4B-Instruct-2507
- Llama-3.1-70B-Instruct vs Phi-3.5-MoE-instruct
- Phi-3.5-MoE-instruct vs Qwen3-Coder-Next
- DeepSeek-Coder-V2-Lite-Instruct vs phi-4
- phi-2 vs Qwen2.5-0.5B-Instruct
- Phi-3-mini-4k-instruct vs Qwen3-4B-Base
- phi-4 vs Qwen3-14B-Base
- Phi-3.5-MoE-instruct vs Qwen2.5-72B-Instruct
- Llama-3.2-3B-Instruct vs Phi-3-mini-4k-instruct
- phi-2 vs Qwen3-1.7B
Best models by task
Phi models appear on these ranked task pages.
Sources
- https://huggingface.co/microsoft/phi-1
- https://huggingface.co/microsoft/phi-1_5
- https://huggingface.co/microsoft/phi-2
- https://huggingface.co/microsoft/Phi-3-mini-4k-instruct
- https://huggingface.co/microsoft/Phi-3-mini-128k-instruct
- https://huggingface.co/microsoft/Phi-3-small-8k-instruct
- https://huggingface.co/microsoft/Phi-3-medium-128k-instruct
- https://huggingface.co/microsoft/Phi-3.5-mini-instruct
- https://huggingface.co/microsoft/Phi-3.5-MoE-instruct
- https://huggingface.co/microsoft/phi-4
- https://huggingface.co/microsoft/Phi-4-mini-instruct
- https://huggingface.co/microsoft/Phi-4-reasoning
- https://huggingface.co/microsoft/Phi-4-multimodal-instruct
- https://huggingface.co/microsoft/Phi-mini-MoE-instruct
- https://huggingface.co/microsoft/Phi-tiny-MoE-instruct
- https://huggingface.co/api/models?author=microsoft&search=phi&sort=createdAt&direction=-1&limit=15
Facts in the text above were read from these pages and are the publisher's own statements, not benchmarks run by Aquanode. Last reviewed 2026-10-07.