Model Follower.

Up to date

Rebuilt
2026-10-25 06:54 UTC
Release radar
117 on watch
Hub snapshot
2026-10-25
Re-checked
daily · every number sourced and dated

Models › phi-4 family

Phi-4-reasoning-plus

Hugging Face ↗ Follow changes

microsoft MIT out 18 months ago

14.7B parameters, text in, text out, 32K tokens of context.

Verified 2026-10-25 Rank #628 in LLMs · score 73.4 method v1 Sources 3 Downloads 30d 4.5k · #554 of 619 tracked

Can I run it?8.8GB at INT4Estimatefits a 16 GB cardmethod
Can I use it?MITPermissivelicence file2026-10-24
Do people use it?4.5kdownloads, 30 days#554 of 619 we track · below the median+2 likes, downloads −32% this week352 likesHugging Face2026-10-24

Worth reading

Picked by rule from 17 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 291 HN pts

Phi 4 available on Ollama

Phi-4 is a 14B parameter open model designed for memory/compute constrained environments and latency bound scenarios.

ollama.com 2025-01-09 291 HN pts · 135 comments 220 words about the phi 4 family family

17 best of 17 tracked

The 2 picks above, plus these 15. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. ArticlePhi-4: Microsoft's Newest Small Language Model Specializing in Complex Reasoningtechcommunity.microsoft.com · 2024-12-13 · 439 HN pts · official
  2. ArticlePhi-4 Reasoning Modelsazure.microsoft.com · 2025-05-01 · 131 HN pts · official
  3. ArticlePhi-4 Bug Fixesunsloth.ai · 2025-01-10 · 193 HN pts
  4. ArticleMicrosoft announces Phi-4-multimodal and Phi-4-miniazure.microsoft.com · 2025-02-26 · official
  5. ArticleIntroducing Phi-4-mini-flash-reasoningazure.microsoft.com · 2025-07-09 · official
  6. ArticlePhi-4-mini and Phi-4-multimodaltechcommunity.microsoft.com · 2025-03-01 · official
  7. ArticleMicrosoft Phi-4 reasoning technical report [pdf]microsoft.com · 2025-05-01 · 2 HN pts · official
  8. ArticlePhi-4 Technical Reportsimonwillison.net · 2024-12-16 · 5 HN pts
  9. ArticleMicrosoft debuts Phi-4, new generative AI modeltechcrunch.com · 2024-12-13 · 24 HN pts
  10. PaperPhi-4 Technical Reportarxiv.org · 2024-12-13 · 3 HN pts
  11. ArticleMicrosoft makes powerful Phi-4 model open-source on Hugging Faceventurebeat.com · 2025-01-09 · 3 HN pts
  12. ArticlePhi-4 as LLM-as-Judge rivals specialized evaluator models wo/ finetuningflow-ai.com · 2025-01-22
  13. ArticleMicrosoft's smaller AI model beats the big guys: Meet Phi-4, the efficiency kingventurebeat.com · 2024-12-13
  14. ArticleReviewing Post-Training Techniques: DeepSeek, Qwen-2, and Phi-4brianfitzgerald.xyz · 2025-01-07
  15. ArticleNew for 2025 – more local models in Pieces, including Qwen Coder and Phi-4pieces.app · 2025-01-31

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

No API provider serves it, so its downloads are people running it themselves: 53 quantized builds and 18 Spaces.

  • As a service0no API provider listed
  • On own hardware53quantized builds
  • Built on1610 fine-tunes · 5 adapters · 1 merges
  • In apps18Hugging Face Spaces

Top Spaces running it: VIDraft/phi-4-reasoning-plus ♥ 24 · NeoPy/phi-4-reasoning-plus ♥ 1

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 16 GB card at INT4

About 8.8 GB at INT4, 18 GB at FP8, 35 GB at BF16: weights only, so leave room for the KV cache.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 8.8 GB of weights (estimate; KV cache and activations extra)INT48.8 GBFP8: about 18 GB of weights (estimate; KV cache and activations extra)FP818 GBBF16: about 35 GB of weights (estimate; KV cache and activations extra)BF1635 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 8.8 GB of weights (estimate; KV cache and activations extra)INT48.8 GBFP8: about 18 GB of weights (estimate; KV cache and activations extra)FP818 GBBF16: about 35 GB of weights (estimate; KV cache and activations extra)BF1635 GB

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

Most coverage came in the first two weeks

Teal is the weights (2025-04-17), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.

JanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOct2025-11-24: 3 repo commits2025-06-13: 3 repo commits2025-05-08: 1 repo commit2025-05-06: 3 repo commits2025-05-02: 1 repo commit2025-04-18: 1 repo commit2025-04-17: 1 repo commitPhi-4-reasoning-vision and the lessons of training a multimodal reasoning model · 2026-03-05Phi-4: Microsoft's Newest Small Language Model Specializing in Complex Reasoning · 2024-12-13Phi-4 Reasoning Models · 2025-05-01Phi-4 Bug Fixes · 2025-01-10Phi 4 available on Ollama · 2025-01-09Microsoft Phi-4 reasoning technical report [pdf] · 2025-05-01Phi-4 Technical Report · 2024-12-16Microsoft debuts Phi-4, new generative AI model · 2024-12-13Phi-4 Technical Report · 2024-12-13Microsoft makes powerful Phi-4 model open-source on Hugging Face · 2025-01-09Weights 2025-04-17 · 2025-04-17Weights 2025-04-17now
JanFebMarAprMayJunJulAugSepOctNovDecJanFebMarAprMayJunJulAugSepOct2025-11-24: 3 repo commits2025-06-13: 3 repo commits2025-05-08: 1 repo commit2025-05-06: 3 repo commits2025-05-02: 1 repo commit2025-04-18: 1 repo commit2025-04-17: 1 repo commitPhi-4-reasoning-vision and the lessons of training a multimodal reasoning model · 2026-03-05Phi-4: Microsoft's Newest Small Language Model Specializing in Complex Reasoning · 2024-12-13Phi-4 Reasoning Models · 2025-05-01Phi-4 Bug Fixes · 2025-01-10Phi 4 available on Ollama · 2025-01-09Microsoft Phi-4 reasoning technical report [pdf] · 2025-05-01Phi-4 Technical Report · 2024-12-16Microsoft debuts Phi-4, new generative AI model · 2024-12-13Phi-4 Technical Report · 2024-12-13Microsoft makes powerful Phi-4 model open-source on Hugging Face · 2025-01-09Weights 2025-04-17 · 2025-04-17Weights 2025-04-17now

What people built on it

69 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

69 derivatives on the Hub, most of them quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

53 quantized (Hugging Face model tree)10 fine-tunes (Hugging Face model tree)1 merge (Hugging Face model tree)5 adapters (Hugging Face model tree)53 quantized10 fine-tunes1 merge5 adapters
53 quantized (Hugging Face model tree)10 fine-tunes (Hugging Face model tree)1 merge (Hugging Face model tree)5 adapters (Hugging Face model tree)53 quantized10 fine-tunes1 merge5 adapters
GGUF unsloth/Phi-4-reasoning-plus-GGUF for llama.cpp, Ollama, LM Studio 309.8k downloads MLX lmstudio-community/Phi-4-reasoning-plus-MLX-4bit for Apple silicon 660.1k downloads NVFP4 nvidia/Phi-4-reasoning-plus-NVFP4 for vLLM, SGLang 745.5k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

What the card says

2 cited sentences

Summary generated, cited

Phi-4-reasoning-plus is a 14.7B parameter dense decoder-only Transformer model finetuned from Phi-4 for reasoning tasks.[1][2] The model is released under an MIT license and supports a context length of 32k tokens.[3][4]

  1. parameters human
  2. “dense decoder-only Transformer model”
  3. license
  4. “Context length | 32k tokens”

Generated 2026-10-07 15:14 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

3 use cases

What it is good at 3 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

22 fields

Listed in LLMs

Facts

Pipeline tagtext-generation
Librarytransformers
ArchitecturesPhi3ForCausalLM
Model typephi3
Dtypebfloat16
Params by dtypeBF16: 14,659,507,200
Size bandL
MoEno
Context length32,768 (config.json max_position_embeddings)
Languagesen
PapersarXiv:2504.21318
LicenseMIT (mit)
Hub snapshot2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · Phi3ForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · Phi3ForCausalLM · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for Phi-4-reasoning-plus: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)35 GB · Single high-end GPU (<80 GB)
FP8 (1 byte/param)18 GB · Single GPU (<24 GB)
INT4 (0.5 bytes/param)8.8 GB · Single GPU (<24 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.

All variants

40 variants

Lineage

Fine-tuned from microsoft/phi-4 (same organization) Family: microsoft/phi-4 69 derivatives on the Hub

Base model: microsoft/phi-4

quantized: 53finetune: 10adapter: 5merge: 1

Top variants (40 shown, ranked by downloads)

Relation
Quant
Size
40 rows
Variant Relation Quant Params Band License Author Likes Downloads
nvidia/Phi-4-reasoning-plus-NVFP4 quantized NVFP4 7.84B M mit nvidia 14 745.5k
lmstudio-community/Phi-4-reasoning-plus-MLX-4bit quantized MLX 14.7B L mit lmstudio-community 1 660.1k
unsloth/Phi-4-reasoning-plus-GGUF quantized GGUF — — mit unsloth 97 309.8k
nvidia/Phi-4-reasoning-plus-FP8 quantized -FP8 14.7B L mit nvidia 7 181.1k
lmstudio-community/Phi-4-reasoning-plus-GGUF quantized GGUF — — mit lmstudio-community 6 150.6k
bartowski/microsoft_Phi-4-reasoning-plus-GGUF quantized GGUF — — mit bartowski 13 48.9k
unsloth/Phi-4-reasoning-plus-unsloth-bnb-4bit quantized -bnb 15.1B L mit unsloth 4 17.2k
mradermacher/Phi-4-reasoning-plus-i1-GGUF quantized GGUF — — mit mradermacher 4 7.6k
second-state/Phi-4-reasoning-plus-GGUF quantized GGUF — — mit second-state 0 7.6k
unsloth/Phi-4-reasoning-plus finetune — 14.7B L mit unsloth 7 6.9k
gaianet/Phi-4-reasoning-plus-GGUF quantized GGUF — — mit gaianet 0 6.6k
unsloth/Phi-4-reasoning-plus-bnb-4bit quantized -bnb 15.1B L mit unsloth 1 6.4k
mlx-community/Phi-4-reasoning-plus-8bit quantized -8bit 14.7B L mit mlx-community 1 6.1k
mlx-community/Phi-4-reasoning-plus-bf16 finetune -bf16 14.7B L mit mlx-community 1 5.5k
mlx-community/Phi-4-reasoning-plus-4bit quantized -4bit 14.7B L mit mlx-community 1 4.2k
steadyflow/Phi-4-reasoning-plus-AWQ quantized AWQ 14.7B L mit steadyflow 1 3.9k
mmnga/Phi-4-reasoning-plus-gguf quantized gguf — — mit mmnga 1 3.8k
mradermacher/Phi-4-reasoning-plus-GGUF quantized GGUF — — mit mradermacher 1 3k
amd/Phi-4-reasoning-plus-w4a16-llmcompressor quantized -w4a16 14.7B L mit amd 0 2.5k
DevQuasar/microsoft.Phi-4-reasoning-plus-GGUF quantized GGUF — — — DevQuasar 0 2k
pbhappliedsystems/phi-4-reasoning-plus-gguf-Q4-K-M quantized gguf — — mit pbhappliedsystems 1 1.4k
amd/Phi-4-reasoning-plus-w4a16-asym-torchao-v0.17.0 quantized -w4a16 8.06B M mit amd 0 1.3k
amd/Phi-4-reasoning-plus-w8a8-llmcompressor quantized -w8a8 14.7B L mit amd 0 1k
SandLogicTechnologies/Phi-4-reasoning-plus-GGUF quantized GGUF — — mit SandLogicTechnologies 0 938
keisuke-miyako/Phi-4-reasoning-plus-gguf-q4_k_m quantized gguf — — — keisuke-miyako 0 925
pbhappliedsystems/phi-4-reasoning-plus-gguf-F16 quantized gguf — — mit pbhappliedsystems 0 788
Jagneshdeveloper/Ekant-14b-small finetune — — — apache-2.0 Jagneshdeveloper 0 759
mlx-community/Phi-4-reasoning-plus-6bit quantized — 14.7B L mit mlx-community 0 678
cortecs/Phi-4-reasoning-plus-FP8-Dynamic quantized -FP8 14.7B L — cortecs 2 515
rubricreward/R3-Phi-4-reasoning-plus-14k adapter — 14.7B L apache-2.0 rubricreward 2 447
peelsa/phi-4-reasoning-plus-ablit finetune — 14.7B L mit peelsa 0 487
OrchardPair/Phi-4-reasoning-plus-4bit quantized -4bit 14.7B L mit OrchardPair 0 392
rubricreward/R3-Phi-4-reasoning-plus-4k finetune — 14.7B L apache-2.0 rubricreward 1 318
NikolayKozloff/Phi-4-reasoning-plus-Q5_K_S-GGUF quantized GGUF — — mit NikolayKozloff 2 266
amd/Phi-4-reasoning-plus-w4a16-tao-symgroup-torchao-v0.17.0 quantized -w4a16 14.9B L mit amd 0 352
amd/Phi-4-reasoning-plus-da8w8-torchao-v0.17.0 quantized — 14.7B L mit amd 0 333
MCES10/Phi-4-reasoning-plus-mlx-fp16 finetune mlx 14.7B L mit MCES10 0 321
prince-canuma/Phi-4-reasoning-Plus-6bit quantized — 14.7B L mit prince-canuma 0 307
jkim96/Phi-4-reasoning-plus-DASHQ-Q2-GGUF quantized GGUF — — mit jkim96 0 299
IWasBorn3D/Phi-4-reasoning-plus-Q6_K-GGUF quantized Q6_K — — mit IWasBorn3D 0 296

Repository changelog

13 commits

Repository changelog 13 commits

DateCommitSHA
2025-11-24 Update data_summary_card.md (#19) 69baf85
2025-11-24 Update README.md (#18) 6367556
2025-11-24 Upload data_summary_card.md (#17) e56ce57
2025-06-13 fix(tokenizer): Sets fim and think tokens to non-special, and removes unk_token. 3670d21
2025-06-13 fix(tokenizer_config): Sets fim and think tokens to non-special, and removes unk_token. f1527a3
2025-06-13 fix(special_tokens_map): Removes unk_token. f580375
2025-05-08 chore(root): Adds top_k information even if 50 is already the default. 0f18025
2025-05-06 Update README.md d8d3b44
2025-05-06 Update README.md 0b70369
2025-05-06 Update README.md 73394c6
2025-05-02 fix(root): Use a better placement for the "Usage" section. (#5) 459ab9f
2025-04-18 chore(root): Adds files. 609962c
2025-04-17 initial commit 28857dc

Articles and coverage

17 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official4

Phi-4-reasoning-vision and the lessons of training a multimodal reasoning model
microsoft.com2026-03-05officialauthorityabout the phi 4 familyHN 93 pts · 11 comments4,049 words
Phi-4-reasoning-vision-15B is a 15 billion parameter open-weight multimodal reasoning model trained with just 200 billion tokens of multimodal data. The model excels at math and science reasoning and understanding user-interfaces, presenting an appealing value relative to popular open-weight models.generated, cited
Phi-4: Microsoft's Newest Small Language Model Specializing in Complex Reasoning
techcommunity.microsoft.com2024-12-13officialauthorityHN front pageabout the phi 4 familyHN 439 pts · 143 comments5,664 words
Microsoft introduces Phi-4, a 14B parameter small language model designed for complex reasoning in math and language tasks. The model achieves performance through the use of high-quality synthetic datasets, curation of high-quality organic data, and post-training innovations. Phi-4 outperforms much larger models, including Gemini Pro 1.5, on math competition problems.written, reviewed, cited
  • Phi-4 is available on Azure AI Foundry and Hugging Face as of the Dec 12, 2024 announcement.
  • Developers can integrate content safety features into applications easily through a single API.
  • The blog post was updated on Jan 08, 2025, indicating potential changes after the initial release.
Phi-4 Reasoning Models
azure.microsoft.com2025-05-01officialauthorityHN front pageabout the phi 4 familyHN 131 pts · 26 comments1,007 words
Microsoft releases Phi-4-reasoning, a 14-billion parameter open-weight model, alongside Phi-4-reasoning-plus and Phi-4-mini-reasoning on Azure AI Foundry and HuggingFace. Phi-4-reasoning-plus utilizes 1.5x more tokens than Phi-4-reasoning to deliver higher accuracy, while both models outperform the 671-billion parameter DeepSeek-R1 on the AIME 2025 test. Optimized versions using ONNX are now available for Snapdragon powered Copilot+ PCs to run efficiently on NPUs.written, reviewed, cited
  • Phi-4-reasoning is trained via supervised fine-tuning of Phi-4 on carefully curated reasoning demonstrations from OpenAI o3-mini.
  • Phi-4-mini-reasoning is fine-tuned with synthetic data generated by Deepseek-R1 model to balance efficiency with advanced reasoning ability.
  • The NPU-optimized Phi Silica variant is designed to be preloaded in memory for blazing fast time to first token responses.
Microsoft Phi-4 reasoning technical report [pdf]
microsoft.com2025-05-01officialauthorityabout the phi 4 familyHN 2 pts · 1 comments

Analysis6

Phi-4 Bug Fixes
unsloth.ai2025-01-10authorityHN front pageabout the phi 4 familyHN 193 pts · 68 comments
Phi 4 available on Ollama
ollama.com2025-01-09HN front pageabout the phi 4 familyHN 291 pts · 135 comments220 words
Phi-4 is a 14B parameter open model designed for memory/compute constrained environments and latency bound scenarios. It supports a context length of 16k tokens and is specifically optimized for reasoning and logic tasks. The model focuses primarily on English, requiring developers to evaluate accuracy and safety before deployment in high-risk scenarios.written, reviewed, cited
Phi-4 Technical Report
simonwillison.net2024-12-16authorityabout the phi 4 familyHN 5 pts · 0 comments740 words
Microsoft debuts Phi-4, new generative AI model
techcrunch.com2024-12-13about the phi 4 familyHN 24 pts · 1 comments
Phi-4 Technical Report
arxiv.org2024-12-13authorityabout the phi 4 familyHN 3 pts · 0 comments
Microsoft makes powerful Phi-4 model open-source on Hugging Face
venturebeat.com2025-01-09about the phi 4 familyHN 3 pts · 1 comments
All coverage tracked for this model (17) — append-only, nothing is ever removed
TitleSourcePublishedFirst seenSignals
Phi-4-reasoning-vision and the lessons of training a multimodal reasoning modelmicrosoft.com official2026-03-052026-10-07hn
Introducing Phi-4-mini-flash-reasoningazure.microsoft.com official2025-07-092026-10-07hn
Microsoft Phi-4 reasoning technical report [pdf]microsoft.com official2025-05-012026-10-07hn
Phi-4 Reasoning Modelsazure.microsoft.com official2025-05-012026-10-07hn
Phi-4-mini and Phi-4-multimodaltechcommunity.microsoft.com official2025-03-012026-10-07hn
Microsoft announces Phi-4-multimodal and Phi-4-miniazure.microsoft.com official2025-02-262026-10-07hn
New for 2025 – more local models in Pieces, including Qwen Coder and Phi-4pieces.app2025-01-312026-10-07hn
Phi-4 as LLM-as-Judge rivals specialized evaluator models wo/ finetuningflow-ai.com2025-01-222026-10-07hn
Phi-4 Bug Fixesunsloth.ai2025-01-102026-10-07hn
Phi 4 available on Ollamaollama.com2025-01-092026-10-07hn
Microsoft makes powerful Phi-4 model open-source on Hugging Faceventurebeat.com2025-01-092026-10-07hn
Reviewing Post-Training Techniques: DeepSeek, Qwen-2, and Phi-4brianfitzgerald.xyz2025-01-072026-10-07hn
Phi-4 Technical Reportsimonwillison.net2024-12-162026-10-07hn
Phi-4 Technical Reportarxiv.org2024-12-132026-10-07hn
Microsoft's smaller AI model beats the big guys: Meet Phi-4, the efficiency kingventurebeat.com2024-12-132026-10-07hn
Microsoft debuts Phi-4, new generative AI modeltechcrunch.com2024-12-132026-10-07hn
Phi-4: Microsoft's Newest Small Language Model Specializing in Complex Reasoningtechcommunity.microsoft.com official2024-12-132026-10-07hn

Papers

1 paper

Papers 1

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

Similar models

3 models

Similar models

  • Qwen3-14B · 14.8B
    same size band (L), same task (text-generation), more providers, permissive license
  • DeepSeek-R1-Distill-Qwen-14B · 14.8B
    same size band (L), same task (text-generation), more providers, permissive license
  • Qwen2.5-14B-Instruct · 14.8B
    same size band (L), same task (text-generation), permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:42 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:14 UTC · Last verified 2026-10-25

Follow Phi-4-reasoning-plus

Watch this model

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.