Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone
Summary
The report introduces phi-3-mini, a 3.8 billion parameter language model trained on 3.3 trillion tokens, which the authors say rivals Mixtral 8x7B and GPT-3.5, for example 69% on MMLU and 8.38 on MT-bench, despite being small enough to deploy on a phone. They also report 7B and 14B models trained for 4.8T tokens, phi-3-small and phi-3-medium, with 75% and 78% on MMLU. A later phi-3.5 series adds mini, MoE, and Vision models. [S1]
Claims
- id: C1 claim: 'phi-3-mini is a 3.8 billion parameter model trained on 3.3 trillion tokens.' source: S1 evidence: "phi-3-mini, a 3.8 billion parameter language model trained on 3.3 trillion tokens"
- id: C2 claim: 'The authors report 69% on MMLU and 8.38 on MT-bench for phi-3-mini.' source: S1 evidence: "phi-3-mini achieves 69% on MMLU and 8.38 on MT-bench"
- id: C3 claim: 'They report phi-3-small and phi-3-medium at 75% and 78% on MMLU after 4.8T tokens.' source: S1 evidence: "respectively 75%, 78% on MMLU, and 8.7, 8.9 on MT-bench"
Verification
- Code available: No official model or training-code URL was named on the abs page or HTML.
- Reproduced: not reproduced (2026-11-12). Code not run.
- Related work: This batch also has the earlier Textbooks Are All You Need (phi-1) note. Live note 2401-04088-mixtral-of-experts.
- Limits: Benchmark figures are author-reported (v4, 30 Aug 2024, which also describes the phi-3.5 series). This is a company technical report, and no model-hosting URL is named in the paper.
- License: CC BY 4.0 on the abs page (2026-11-12; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 22 Apr 2024 (v1); last revised 30 Aug 2024 (v4). Sources: [S1] https://arxiv.org/abs/2404.14219, v4 (30 Aug 2024). checked_on: 2026-11-12.
When to apply
Use when asking what the Phi-3 report claimed for phone-scale models and later 7B, 14B, and phi-3.5 variants.
Facts
- id: F1 claim: 'phi-3-mini has 3.8 billion parameters and was trained on 3.3 trillion tokens.' confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
- id: F2 claim: 'The authors report 69% MMLU and 8.38 MT-bench for phi-3-mini.' confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
- id: F3 claim: 'They also report phi-3-small and phi-3-medium, trained for 4.8T tokens, at 75% and 78% on MMLU.' confidence: medium date: 2026-11-12 sources: [S1] license: CC BY 4.0
Sources
- [S1] arXiv 2404.14219, https://arxiv.org/abs/2404.14219. License: CC BY 4.0. v4 (30 Aug 2024). Accessed 2026-11-12.
License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.
License: CC-BY-4.0 · Summary by iinventedlight (AI-assisted). Source: Marah Abdin and 128 coauthors (listed on the arXiv abs page), Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone (CC BY 4.0).
Cite this
'iinventedlight (AI-assisted)'. 2026. "Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone." iinventedlight Library, iil:note:nbch8z66, version of 2026-11-12. https://iinventedlight.com/library/notes/2404-14219-phi-3/. License: CC-BY-4.0.
@misc{iil_note_nbch8z66,
author = {{'iinventedlight (AI-assisted)'}},
title = {{Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone}},
howpublished = {iinventedlight Library},
year = {2026},
date = {2026-11-12},
url = {https://iinventedlight.com/library/notes/2404-14219-phi-3/},
note = {Permanent ID iil:note:nbch8z66. License: CC-BY-4.0.},
}
Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.
For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/2404-14219-phi-3.md.