Skip to content
AI Atlas
ModelActiveOpen weightsIdentity probable

Phi 4

Microsoftfamily · Phi4huggingface.co/microsoft/phi-4

[Microsoft Research](/microsoft) Phi-4 is designed to perform well in complex reasoning tasks and can operate efficiently in situations with limited memory or where quick responses are needed. At 14 billion...

Open in Graph
data quality72

Updated 6 h ago · first seen 11 Sept 2026

model_01M294WWM26ZA9FAW9BCD0QYCJ

Overview

Identity

Canonical model
Yesidentity confidence: mediumOne row per real model release. Artifacts (checkpoints, quantisations, conversions) and folded evaluation variants point here.
Official checkpoints
official_checkpoints = hf_repo identifiers carried by the model itself; artifacts are separate entities pointing here through canonical_id.
Artifacts
None recordedSeparate entities (checkpoint · quantization · conversion · packaging) pointing to this model through canonical_id.
Provider deployments
1
API aliases
microsoft/phi-4phi-4Identifiers under which providers and evaluators refer to this model.
Folded evaluation variants
0Effort / thinking variants (…-high, …-non-reasoning) are result configurations of this model, not separate models. Their old URLs redirect here.

Openness

Open weights— weights downloadable under MIT; commercial use allowed; redistribution allowed; derivatives allowed; 4 dimensions unknown.

Weights downloadable under a permissive or Creative Commons licence allowing commercial use; code or data may be missing.

  • Weights

    Yes

  • Inference code

    —

  • Training code

    —

  • Training data

    —

  • Dataset

    —

  • Commercial use

    Yes

  • Redistribution

    Yes

  • Derivatives

    Yes

Licence: MIT License (permissive · SPDX MIT · stated as “mit”)

dimensions marked null are unknown, not false

Key facts

Release date

Source:Hugging Face Hub (public pages, model cards, papers)T2observed 11 h agomedium

Knowledge cutoff

Source:OpenRouter public model & pricing listingT2observed 16 h agomedium

Model card
huggingface.co/microsoft/phi-4

Source:Hugging Face Hub (public pages, model cards, papers)T2observed 16 h agomedium

Openrouter id

Source:OpenRouter public model & pricing listingT2observed 16 h agomedium

Architecture

Architecture

Source:Hugging Face Hub (public pages, model cards, papers)T2observed 15 h agomedium

Model type

Source:Hugging Face Hub (public pages, model cards, papers)T2observed 15 h agomedium

Parameters

Source:Hugging Face Hub (public pages, model cards, papers)T2observed 16 h agomedium

Weights dtype

Source:Hugging Face Hub (public pages, model cards, papers)T2observed 15 h agomedium

File size

Source:Hugging Face Hub (public pages, model cards, papers)T2observed 15 h agomedium

Library name

Source:Hugging Face Hub (public pages, model cards, papers)T2observed 15 h agomedium

Pipeline tag

Source:Hugging Face Hub (public pages, model cards, papers)T2observed 16 h agomedium

Hugging Face repo

Source:OpenRouter public model & pricing listingT2observed 16 h agomedium

Capabilities

Modalities

Modalities
text
Input
text
Output
text

Capabilities

  • Tool calling

    No

    OpenRouter public model & pricing listing · T2

  • Structured output

    Yes

    OpenRouter public model & pricing listing · T2

  • Reasoning

    No

    Artificial Analysis · T2

  • Vision

    Unavailable

  • Audio

    Unavailable

  • Fine-tuning available

    Unavailable

Context window

Source:OpenRouter public model & pricing listingT2observed 13 h agomedium

Max output

Source:OpenRouter public model & pricing listingT2observed 16 h agomedium

Knowledge cutoff

Source:OpenRouter public model & pricing listingT2observed 16 h agomedium

Languages

Source:Hugging Face Hub (public pages, model cards, papers)T2observed 15 h agomedium

Benchmarks12

Compare with another model →

Comparable same task and conditions · Partially comparable same task, conditions differ (effort, temperature, judge) · Not comparable different variant or metric

Benchmark results grouped by comparability group
Benchmark · groupBest scoreTrustConfigurationResultsvs leaderEvaluatedSource
Terminal-Benchagentic · accuracy · variant=hard · evaluator=Artificial AnalysisIndependentevaluatorArtificial Analysisvarianthardreasoningoffconditions differ across rows → partially comparable2−62.1 ptvs gpt-5.6-solobs. 12 Sept 2026artificialanalysis.aiT2
τ²-benchagentic · pass^1 · variant=Telecom · evaluator=Artificial AnalysisIndependentevaluatorArtificial AnalysisvariantTelecomreasoningoffconditions differ across rows → partially comparable1non-primary groupobs. 12 Sept 2026artificialanalysis.aiT2
τ²-benchagentic · pass^1 · evaluator=Artificial AnalysisIndependentevaluatorArtificial Analysis1−99.1 ptvs Z.ai GLM 5.2obs. 11 Sept 2026artificialanalysis.aiT2
Artificial Analysis Intelligence Indexcomposite · indexIndependentreasoningoffversion4.3conditions differ across rows → partially comparable2−47.4vs Claude Fable 5.1obs. 12 Sept 2026artificialanalysis.aiT2
IFBenchinstruction-following · accuracy · evaluator=Artificial AnalysisIndependentevaluatorArtificial Analysisreasoningoffconditions differ across rows → partially comparable2−59.8 ptvs Grok 4.3obs. 12 Sept 2026artificialanalysis.aiT2
Humanity's Last Examknowledge · accuracy · evaluator=Artificial AnalysisIndependentevaluatorArtificial Analysisreasoningoffconditions differ across rows → partially comparable2−55.4 ptvs Claude Fable 5.1obs. 12 Sept 2026artificialanalysis.aiT2
GPQA Diamondreasoning · accuracy · variant=Diamond · evaluator=Artificial AnalysisIndependentevaluatorArtificial AnalysisvariantDiamondreasoningoffconditions differ across rows → partially comparable1non-primary groupobs. 12 Sept 2026artificialanalysis.aiT2
GPQA Diamondreasoning · accuracy · variant=GPQA Diamond · evaluator=Artificial AnalysisIndependentevaluatorArtificial AnalysisvariantGPQA Diamond1−38.8 ptvs gpt-6-astraobs. 11 Sept 2026artificialanalysis.aiT2

Current rows only, grouped by benchmark → canonical metric → comparability group (task configuration). Effort variants folded into this model appear as rows of the same group. 12 current rows in total. “vs leader” compares with the current leader of the benchmark's primary group only; other groups are not directly comparable. Comparability rules →

Providers & Pricing1

All offers in the price terminal →
Provider deployments, cheapest output first
ProviderContextInput / 1MOutput / 1MStatusObservedSource
OpenRoutercheapest outputmicrosoft/phi-416.4Kout 14.8Kactive2 h agosince 11 Sept 2026openrouter.aiT2

USD per 1M tokens as published by each provider; native units (per-request fees, flex/priority tiers) are kept verbatim. Rows are append-only — every price change is kept in the history below. Cost of a workload →

Price history

Step lines per provider; amber markers are recorded changes. Click a marker or a row for the evidence behind that price.

Output price · USD / 1M tokens 1 provider

Output price history of Phi 4$0$0.05$0.10$0.15$0.20Sept 26Sept 26Sept 26Sept 26Sept 26OpenRouter: first observed → $0.14 · 11 Sept 2026
  • OpenRouter
  • OpenRouterfirst observed $0.1411 Sept 2026

Input price · USD / 1M tokens 1 provider

Input price history of Phi 4$0$0.02$0.04$0.06$0.08$0.10Sept 26Sept 26Sept 26Sept 26Sept 26OpenRouter: first observed → $0.07 · 11 Sept 2026
  • OpenRouter
  • OpenRouterfirst observed $0.0711 Sept 2026

Hardware fit37

Estimated

37 of 37 device × quantization combinations fit.

Run locally: your machine →
Estimated hardware fit
HardwareQuantizationDevice memoryEst. memoryFits
Apple M3 Ultra4bit—8.9 GB est.Yes
Mac Studio (Apple M5 Ultra)4bit—8.9 GB est.Yes
Apple M2 Ultra4bit—8.9 GB est.Yes
Apple M1 Ultra4bit—8.9 GB est.Yes
Apple M3 Max4bit—8.9 GB est.Yes
Apple M4 Max4bit—8.9 GB est.Yes
Mac Studio (Apple M5 Max)4bit—8.9 GB est.Yes
MacBook Pro (Apple M5 Max)4bit—8.9 GB est.Yes
Apple M2 Max4bit—8.9 GB est.Yes
Apple M1 Max4bit—8.9 GB est.Yes
Apple M4 Pro4bit—8.9 GB est.Yes
Mac mini (Apple M5 Pro)4bit—8.9 GB est.Yes
MacBook Pro (Apple M5 Pro)4bit—8.9 GB est.Yes
Apple M3 Pro4bit—8.9 GB est.Yes
Apple M1 Pro4bit—8.9 GB est.Yes
Apple M2 Pro4bit—8.9 GB est.Yes
Apple M44bit—8.9 GB est.Yes
iMac (Apple M4)4bit—8.9 GB est.Yes
Mac mini (Apple M6)4bit—8.9 GB est.Yes
MacBook Air (Apple M5)4bit—8.9 GB est.Yes
MacBook Pro (Apple M5)4bit—8.9 GB est.Yes
Apple M24bit—8.9 GB est.Yes
Apple M34bit—8.9 GB est.Yes
Apple M14bit—8.9 GB est.Yes
NVIDIA GeForce RTX 30904bit24 GB8.9 GB est.Yes
NVIDIA GeForce RTX 40904bit24 GB8.9 GB est.Yes
NVIDIA GeForce RTX 50904bit32 GB8.9 GB est.Yes
NVIDIA A100 80GB4bit80 GB8.9 GB est.Yes
NVIDIA H100 SXM4bit80 GB8.9 GB est.Yes
NVIDIA H100 NVL4bit94 GB8.9 GB est.Yes
NVIDIA DGX Spark4bit128 GB8.9 GB est.Yes
NVIDIA H2004bit141 GB8.9 GB est.Yes
NVIDIA H200 NVL4bit141 GB8.9 GB est.Yes
NVIDIA B2004bit180 GB8.9 GB est.Yes
AMD Instinct MI300X4bit192 GB8.9 GB est.Yes
AMD Instinct MI325X4bit256 GB8.9 GB est.Yes
NVIDIA DGX B2004bit1,440 GB8.9 GB est.Yes
Assumptions (7)
  • Estimated, not measured: weights = parameters × bytes/param × 1.15 runtime overhead (or the observed artifact file size when one is recorded).
  • bytes/param: 4bit = 0.5, 8bit = 1.0, fp16 = 2.0 (uniform quantization, no per-layer exceptions).
  • KV cache: 2 × layers × kv_heads × head_dim × 2 bytes × context × batch when the architecture is known; otherwise 0.5 GB per 8 192 tokens (× batch), independent of architecture (GQA/MLA models need less).
  • A model 'fits' when the estimate is at most the device memory minus 2 GB reserved for the OS and framework.
  • Mixture-of-experts models are estimated on total parameters (all experts must be resident); active parameters are ignored.
  • Device memory uses the largest configuration when several are listed (e.g. Apple silicon tiers).
  • Multi-GPU: device memories are summed; interconnect bandwidth, tensor-parallel replication and pipeline bubbles are not modelled.

Versions & Artifacts0

Version history

Context window2 changes

11 Sept 2026→11 Sept 2026→12 Sept 2026current

Knowledge cutofffirst observation only

11 Sept 2026current

Licensefirst observation only

11 Sept 2026current

Max outputfirst observation only

11 Sept 2026current

Opennessfirst observation only

11 Sept 2026current

Parametersfirst observation only

11 Sept 2026current

Each hop is a claim: click a value for its source, tier and observation time. Nothing is overwritten — a new observation closes the previous claim.

Artifacts 0

No artifact (checkpoint, quantisation, conversion or packaging) points to this model yet.

Papers1

Timeline9

Full timeline →
  • Benchmark resultModelPhi 4Microsoft

    Phi 4 scores 3.79% on Terminal-Bench

    artificial_analysis
    artificialanalysis.ai
  • Benchmark resultModelPhi 4Microsoft

    Phi 4 scores 0% on τ²-bench

    artificial_analysis
    artificialanalysis.ai
  • Benchmark resultModelPhi 4Microsoft

    Phi 4 scores 23.54% on IFBench

    artificial_analysis
    artificialanalysis.ai
  • Benchmark resultModelPhi 4Microsoft

    Phi 4 scores 3.76% on Humanity's Last Exam

    artificial_analysis
    artificialanalysis.ai
  • Benchmark resultModelPhi 4Microsoft

    Phi 4 scores 57.47% on GPQA Diamond

    artificial_analysis
    artificialanalysis.ai
  • Benchmark resultModelPhi 4Microsoft

    Phi 4 scores 5.92 on Artificial Analysis Intelligence Index

    artificial_analysis
    artificialanalysis.ai
  • Context window changedModelPhi 4Microsoft

    Phi 4: context length changed from 16000 to 16384

    Context window16K tokens→16.4K tokensopenrouter
    openrouter.ai
  • Release date changedModelPhi 4Microsoft

    Phi 4: release date changed from 2024-12-11 to 2025-01-10

    Release date11 Dec 2024→10 Jan 2025openrouter
    openrouter.ai
  • Release date changedModelPhi 4Microsoft

    Phi 4: release date changed from 2025-01-10 to 2024-12-11

    Release date10 Jan 2025→11 Dec 2024huggingface
    huggingface.co

Change history

Temporal, append-only claims: a new observation closes the previous claim instead of overwriting it. Rewind the record with the as-of picker.
1 claims · 1 propertiesShow all properties

Aa median output tokens per secondmetric.aa_median_output_tokens_per_second1

Claim history for Aa median output tokens per second
ValueValid from → toStatusSourceConfidenceExtractor
40.7 → currentcurrentArtificial AnalysisT2mediumdeterministic

Claims are temporal and append-only: a new observation closes the previous claim (valid_to) instead of overwriting it. Conflicting claims from different sources are kept side by side and flagged — never averaged. Methodology →

Provenance

Attributed facts

43

Source tiers

T243

Freshest observation

6 h ago

Conflicts

None

Source documents 5

Source documents
SourceDocumentTypeTierLast observedSnapshots
OpenRouter public model & pricing listingopenrouter.ai/api/v1/models catalogueT2· Quality secondary2 h ago11
Artificial Analysisartificialanalysis.ai/leaderboards/models leaderboardT2· Quality secondary6 h ago2
Hugging Face Hub (public pages, model cards, papers)huggingface.co/models?author=microsoft&p=0&sort=downloads listingT2· Quality secondary6 h ago7
Hugging Face Hub (public pages, model cards, papers)huggingface.co/microsoft/phi-4/raw/main/README.md model_cardT2· Quality secondary9 h ago1
Hugging Face Hub (public pages, model cards, papers)huggingface.co/microsoft/phi-4 model_pageT2· Quality secondary10 h ago5

Tier 1 = official/primary, 2 = quality secondary, 3 = community, 4 = unverified. Every snapshot is archived; see all sources and the methodology.

Data quality (72/100) measures how well AI Atlas knows this entity — completeness, primary-source ratio, freshness, conflicts — never how good the model is. Methodology →