Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone

Capability at sub-10B scale is dominated by parameter count and data volume, which makes such models too expensive to run on-device even though on-device deployment is where the latency and cost pressure actually is.

Not an inference-time mechanism but a TRAINING-RECIPE argument: phi-3-mini is 3.8B parameters trained on 3.3 TRILLION tokens on a scaled-up version of the phi-2 dataset composed of heavily filtered public web data plus heavy SYNTHETIC data. The claim is that data QUALITY filtering and synthetic data generation substitute for scale. Also provides parameter-scaling results (7B/14B at 4.8T tokens) and a phi-3.5 series including phi-3.5-MoE (16 x 3.8B, 6.6B active parameters) and phi-3.5-Vision.

Not a speedup claim. Capability at a given size: phi-3-mini achieves 69% MMLU and 8.38 MT-bench, rivaling Mixtral 8x7B and GPT-3.5, while being small enough to deploy on a phone. phi-3.5-MoE (6.6B active) is on par with Gemini-1.5-Flash and GPT-4o-mini. The INFERENCE-relevant claim is the on-device feasibility itself, which the abstract states qualitatively and does not quantify with a tok/s figure - so no mobile throughput number is recorded here because none is given.

ADOPTION: in-production - weights are open (open_weights true) and phi-3-mini is genuinely deployed on phones and other constrained hardware, so this is one of the few records here where the artifact, not just the idea, is in production. VENUE NOTE: a self-described technical report with no venue; arXiv comment says only '24 pages', so venue = arxiv-preprint, year = ArXiv v1 (Apr 2024). INCLUDED AS THE SMALL-MODEL LINEAGE ANCHOR alongside MobileLLM and Mistral 7B because it makes the explicit trade this repo cares about: a smaller model at equal capability is a strictly cheaper inference, and at 3.8B the whole model fits a phone, which is the strongest possible version of the argument. The 3.3T-token training cost is the offsetting expense and is not itself quantified in energy terms here - for the carbon accounting, route via pap-eff-carbon-emissions-large-neural-training rather than expecting it in this record. DANGLING REF STILL OPEN (2026-10-04 paper-read pass; supersedes the earlier 'LEFT UNRESOLVED' note, which reached the same conclusion but only from this record's own text - this time the PAPER itself was read). No A100 and no training hardware named anywhere in the paper. Its only hardware statements are about the deployed phone: "we tested the quantized model by deploying phi-3-mini on iPhone 14 with A16 Bionic chip running natively on-device", and "it only occupies ~1.8GB of memory". Source: https://arxiv.org/pdf/2404.14219 . Left dangling: the paper states nothing about the GPUs used to train the model. AFFILIATIONS FILLED 2026-10-04 (paper-verification pass): from the Semantic Scholar graph API (paper/arXiv:2404.14219), which is the only machine-readable source that carries author affiliations — the arXiv Atom feed returns author names only, and no ORCID is exposed by either API. One known upstream typo corrected: "Tel Aviv Univesity" -> "Tel Aviv University". S2 affiliation coverage is partial (120 of 249 corpus records carry any), so these lists may be incomplete rather than exhaustive. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'unknown'. `venue` is 'arxiv-preprint', which means no peer-reviewed venue has been established for this work as of the verification date, so the track genuinely is not known. 'unknown' is used here rather than leaving the field null or guessing 'main': the schema separates 'unknown' from 'main' precisely so that an unprereviewed preprint is not silently counted as main-track work. This annotation is what lets a downstream filter distinguish 'peer reviewed, main track' from 'not peer reviewed' without inferring it from the venue string.