Phi-tiny-MoE-instruct

microsoft MIT out 16 months ago

3.76B parameters (2 of 16 experts per token), text in, text out, 4K tokens of context.

Verified 2026-10-25 Rank #771 in LLMs · score 71.2 method v1 Sources 3 Downloads 30d 31k · #420 of 619 tracked

Can I run it?2.3GB at INT4Estimatefits a 16 GB cardmethod

Can I use it?MITPermissivelicence file2026-10-24

Do people use it?31kdownloads, 30 days#420 of 619 we track · below the median+1 likes, downloads −21% this week48 likesHugging Face2026-10-24

Worth reading

Picked by rule from 1 tracked piece: the lab's own post, the most shared, a field report, the latest.

Latest · 22 months ago

Phi-4: Microsoft's Newest Small Language Model Specializing in Complex Reasoning

Microsoft introduces Phi-4, a 14B parameter small language model designed for complex reasoning in math and language tasks.

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.

techcommunity.microsoft.com 2024-12-13 439 HN pts · 143 comments 5.7k words about the Phi family family

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: 2 quantized builds and 2 Spaces.

As a service0no API provider listed
On own hardware2quantized builds
Built on105 fine-tunes · 5 adapters · 0 merges
In apps2Hugging Face Spaces

Will it fit?

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

Fits a 16 GB card at INT4

About 2.3 GB at INT4, 4.5 GB at FP8, 9.0 GB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.

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

The story so far

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

1 dated article and 10 repo commits so far

Teal is the weights (2025-06-23), amber triangles are 1 dated article (each links to it), grey ticks are repo commits.

What people built on it

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

12 derivatives on the Hub, led by 5 fine-tunes

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

GGUF tripathyShaswata/Phi-tiny-MoE-instruct-GGUF for llama.cpp, Ollama, LM Studio 1.5k downloads GPTQ ModelCloud/Phi-tiny-MoE-instruct-GPTQ-W4-MixedGroup-G32-G128 for vLLM, SGLang 139 downloads

All variants →

Reference

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

Benchmarks

7 rows

Benchmarks 7 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (7)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.

Benchmark	Value	Setting / notes	Reported by	Date
author-reported ARC Challenge · Chat and follow instructions	76.37	Phi-tiny-MoE	model author	2026-10-07
author-reported BBH · Reason through hard problems	45.58	Phi-tiny-MoE	model author	2026-10-07
author-reported GSM8K · Reason through hard problems	78.47	Phi-tiny-MoE	model author	2026-10-07
author-reported HumanEval · Write code with agents	58.5	Phi-tiny-MoE	model author	2026-10-07
author-reported MMLU · Reason through hard problems	60.83	Phi-tiny-MoE	model author	2026-10-07
author-reported MMLU-Pro · Reason through hard problems	36.34	Phi-tiny-MoE	model author	2026-10-07
author-reported MT-Bench · Chat and follow instructions	7.05	Phi-tiny-MoE	model author	2026-10-07

Use-case evidence

4 use cases

What it is good at 4 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.

Chat and follow instructionsauthor-reportedAuthor reports Arc-C (chat) 76.37 Author reports MT-bench 7.05 decision-model estimate p=1.00 +3 more
Reason through hard problemsauthor-reportedAuthor reports MMLU 60.83 Author reports MMLU pro 36.34 Author reports BBH 45.58 +1 more
Write code with agentsauthor-reportedAuthor reports Human-eval 58.50
Run on a laptop or phonestructuraldecision-model estimate p=0.94 name: tiny fits: single-gpu

Facts

24 fields

Listed in LLMs · MoE

Facts

Phi-tiny-MoE-instruct has 3.76B parameters. (safetensors metadata, observed 2026-10-24)
Phi-tiny-MoE-instruct is a mixture-of-experts model with 16 experts, routing 2 per token. (config.json, observed 2026-10-24)
Context length is 4,096 tokens (4K). (config.json max_position_embeddings, observed 2026-10-24)
License: MIT (friction: none). Permissive. (license file, observed 2026-10-24)
Created 2025-06-23, last modified 2025-12-10. (HF API, observed 2026-10-24)

Pipeline tag	text-generation
Library	transformers
Architectures	PhiMoEForCausalLM
Model type	phimoe
Dtype	bfloat16
Params by dtype	BF16: 3,755,220,288
Size band	M
MoE	yes
Experts	16
Experts per token	2
Context length	4,096 (config.json max_position_embeddings)
Languages	en
Papers	arXiv:2506.18349, arXiv:2404.14219, arXiv:2409.12136
License	MIT (mit)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · PhiMoEForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · PhiMoEForCausalLM · 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-tiny-MoE-instruct: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	9.0 GB · Single GPU (<24 GB)
FP8 (1 byte/param)	4.5 GB · Single GPU (<24 GB)
INT4 (0.5 bytes/param)	2.3 GB · Edge (<4 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.

All variants

12 variants

Lineage

Original model Family: microsoft/Phi 12 derivatives on the Hub

finetune: 5quantized: 2adapter: 5

Top variants (12 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
yujiepan/phi-moe-tiny-random	finetune	—	2.50M	S	—	yujiepan	1	3k
tiny-random/phi-moe	finetune	—	2.50M	S	—	tiny-random	0	2.7k
tripathyShaswata/Phi-tiny-MoE-instruct-GGUF	quantized	GGUF	—	—	mit	tripathyShaswata	1	1.5k
runanywhere/phi_tiny_moe_HNPU	finetune	—	—	—	mit	runanywhere	0	393
sayande/AgriQA-Phi-MoE-1B	adapter	—	—	—	mit	sayande	0	255
rkumar1999/phi-tiny-moe-math-lean-sft	finetune	—	3.76B	M	—	rkumar1999	0	180
rkumar1999/phi-tiny-moe-lean-sft	finetune	—	3.76B	M	—	rkumar1999	0	177
ModelCloud/Phi-tiny-MoE-instruct-GPTQ-W4-MixedGroup-G32-G128	quantized	GPTQ	3.76B	M	mit	ModelCloud	0	139
ddidacus/phi-tiny-moe-cache-reward	adapter	—	—	—	—	ddidacus	0	30
ddidacus/phi-tiny-moe-controller	adapter	—	—	—	—	ddidacus	0	29
ddidacus/phi-tiny-moe-melinoe	adapter	—	—	—	—	ddidacus	0	25
ddidacus/phi-tiny-moe-prompt-conditioned	adapter	—	—	—	—	ddidacus	0	24

Repository changelog

10 commits

Repository changelog 10 commits

Date	Commit	SHA
2025-12-10	Update README.md (#6)	2fe50e8
2025-12-10	Upload data_summary_card.md (#5)	47e2c49
2025-07-23	Update README.md	1211092
2025-07-20	Add library_name and link to code (#1)	df48363
2025-07-17	Fix minor error (#2)	98ee0fa
2025-06-24	Upload folder using huggingface_hub	0c3e9df
2025-06-24	Update README.md	11d29b2
2025-06-23	Update README.md	b85df31
2025-06-23	Upload folder using huggingface_hub	acd1d1b
2025-06-23	initial commit	413c01e

Articles and coverage

1 piece

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.

Official1

Phi-4: Microsoft's Newest Small Language Model Specializing in Complex Reasoning

techcommunity.microsoft.com2024-12-13officialauthorityHN front pageabout the Phi 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.

Papers

3 papers

Papers 3

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).

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

Marah Abdin, Jyoti Aneja, Hany Awadalla +126 · 2024-04-22 · PDF · HF paper page (262 upvotes)
GRIN: GRadient-INformed MoE this model

Liyuan Liu, Young Jin Kim, Shuohang Wang +14 · 2024-09-18 · PDF · HF paper page (17 upvotes)
SlimMoE: Structured Compression of Large MoE Models via Expert Slimming and Distillation this model

Zichong Li, Chen Liang, Zixuan Zhang +4 · 2025-06-23 · PDF · HF paper page (13 upvotes)

Similar models

3 models

Similar models

granite-4.0-h-tiny · 6.94B
same size band (M), also MoE, same task (text-generation), permissive license
OLMoE-1B-7B-0924 · 6.92B
same size band (M), also MoE, same task (text-generation), permissive license
Qwen3-4B · 4.02B
same size band (M), same task (text-generation), more providers, 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-tiny-MoE-instruct

Watch this model

/feeds/models/microsoft--Phi-tiny-MoE-instruct.xml(RSS 2.0) · every change to Phi-tiny-MoE-instruct
/feeds/families/microsoft--Phi.xml(RSS 2.0) · every change across the Phi family

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