Up to date
microsoft MIT out 22 months ago
14.7B parameters, text in, text out, 16K tokens of context. 2 providers serve it from $0.056 per million input tokens.
Verified 2026-10-25 Rank #56 in LLMs · score 106.1 method v1 Sources 4 Downloads 30d 364.8k · #192 of 619 tracked
Changed this week: Cheapest input price fell 20%: $0.070 → $0.056 per 1M (OpenRouter).2026-10-24 · 1 new quantized build on the Hub (169 in all).2026-10-24
Picked by rule from 27 tracked pieces: the lab's own post, the most shared, a field report, the latest.
Phi-4 is available on Azure AI Foundry and Hugging Face as of the Dec 12, 2024 announcement.
Microsoft introduces Phi-4, a 14B parameter small language model designed for complex reasoning in math and language tasks.
Phi-4 is a 14B parameter open model designed for memory/compute constrained environments and latency bound scenarios.
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 3 picks above, plus these 22. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.
Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.
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.
Served by 2 API providers; heavily run locally (169 quantized builds); a big base for builders (337 fine-tunes, adapters and merges).
Top Spaces running it: pliny-the-prompter/obliteratus ♥ 621 · reach-vb/2024-ai-timeline ♥ 467 · eswardivi/phi-4 ♥ 133 · Agents-MCP-Hackathon/ShallowCodeResearch ♥ 51
Weights only, an estimate: leave room for the KV cache. Method
About 8.8 GB at INT4, 18 GB at FP8, 35 GB at BF16: weights only, so leave room for the KV cache.
Estimate params × bytes per param × 1.2 · how we estimate
Third-party measurements only; author-reported numbers stay in Reference, marked as such.
28.79. Each tick is a model on this site with a measured BFCL row; author-reported numbers stay in Reference. Percent (overall accuracy). Prompt
Source · leaderboard · 2026-10-25
Weights, every dated article, and repo commits on one rail.
Teal is the weights (2024-12-11), amber triangles are 8 dated articles (each links to it), grey ticks are repo commits.
506 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.
Everything else, folded: the full tables, every source.
| Benchmark | Value | Setting / notes | Reported by | Date |
|---|---|---|---|---|
| measured BFCL · Call tools and APIs | 28.79 percent (overall accuracy) | Prompt5 sub-scores
| leaderboard | 2026-10-25 |
| author-reported DROP | 75.5 | phi-4 (14B) | model author | 2026-10-07 |
| author-reported GPQA Diamond · Reason through hard problems | 56.1 | phi-4 (14B) | model author | 2026-10-07 |
| author-reported HumanEval · Write code with agents | 82.6 | phi-4 (14B) | model author | 2026-10-07 |
| author-reported MATH-500 · Reason through hard problems | 80.4 | phi-4 (14B) | model author | 2026-10-07 |
| author-reported MGSM · Work across languages | 80.6 | phi-4 (14B) | model author | 2026-10-07 |
| author-reported MMLU · Reason through hard problems | 84.8 | phi-4 (14B) | model author | 2026-10-07 |
| author-reported SimpleQA · Reason through hard problems | 3 | phi-4 (14B) | model author | 2026-10-07 |
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.
Listed in LLMs
| Pipeline tag | text-generation |
|---|---|
| Library | transformers |
| Architectures | Phi3ForCausalLM |
| Model type | phi3 |
| Dtype | bfloat16 |
| Params by dtype | BF16: 14,659,507,200 |
| Size band | L |
| MoE | no |
| Context length | 16,384 (config.json max_position_embeddings) |
| Languages | en |
| Papers | arXiv:2412.08905 |
| License | MIT (mit) |
| Hub snapshot | 2026-10-25 |
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.
Weights-only memory for phi-4: 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.
quantized: 169finetune: 127adapter: 208merge: 2
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| MaziyarPanahi/phi-4-GGUF | quantized | GGUF | — | — | — | MaziyarPanahi | 10 | 5.5M |
| mlx-community/phi-4-4bit | quantized | -4bit | 14.7B | L | mit | mlx-community | 23 | 878.4k |
| unsloth/phi-4-unsloth-bnb-4bit | quantized | -bnb | 15.1B | L | mit | unsloth | 65 | 783.5k |
| unsloth/phi-4 | finetune | — | 14.7B | L | mit | unsloth | 91 | 677.4k |
| lmstudio-community/phi-4-GGUF | quantized | GGUF | — | — | mit | lmstudio-community | 56 | 412.6k |
| stelterlab/phi-4-AWQ | quantized | AWQ | 14.7B | L | mit | stelterlab | 5 | 303.4k |
| unsloth/phi-4-GGUF | quantized | GGUF | — | — | mit | unsloth | 194 | 246.9k |
| bartowski/phi-4-GGUF | quantized | GGUF | — | — | mit | bartowski | 65 | 238k |
| RedHatAI/phi-4-FP8-dynamic | quantized | -FP8 | 14.7B | L | mit | RedHatAI | 3 | 184.3k |
| unsloth/phi-4-bnb-4bit | quantized | -bnb | 15.1B | L | mit | unsloth | 17 | 163.7k |
| RedHatAI/phi-4-quantized.w4a16 | quantized | -quant | 14.7B | L | mit | RedHatAI | 6 | 144.5k |
| cortecs/phi-4-FP8-Dynamic | quantized | -FP8 | 14.7B | L | — | cortecs | 5 | 81.9k |
| mlx-community/phi-4-8bit | quantized | -8bit | 14.7B | L | mit | mlx-community | 14 | 55.5k |
| RedHatAI/phi-4-quantized.w8a8 | quantized | -quant | 14.7B | L | mit | RedHatAI | 5 | 50.5k |
| curiousmind147/microsoft-phi-4-AWQ-4bit-GEMM | quantized | AWQ | 14.7B | L | mit | curiousmind147 | 2 | 49.2k |
| mlx-community/phi-4-bf16 | finetune | -bf16 | 14.7B | L | mit | mlx-community | 2 | 43k |
| dnotitia/DNA-R1 | finetune | — | 14.7B | L | cc-by-nc-4.0 | dnotitia | 41 | 16.1k |
| jakiAJK/microsoft-phi-4_GPTQ-int4 | quantized | GPTQ | 14.7B | L | — | jakiAJK | 2 | 18k |
| mradermacher/phi-4-i1-GGUF | quantized | GGUF | — | — | mit | mradermacher | 3 | 14.7k |
| mradermacher/phi-4-GGUF | quantized | GGUF | — | — | mit | mradermacher | 1 | 14.8k |
| Mungert/Phi-4-reasoning-plus-GGUF | quantized | GGUF | — | — | mit | Mungert | 4 | 13k |
| mmnga/phi-4-gguf | quantized | gguf | — | — | mit | mmnga | 0 | 11.1k |
| second-state/phi-4-GGUF | quantized | GGUF | — | — | mit | second-state | 0 | 9.5k |
| WestlakeNLP/DeepReviewer-14B | finetune | — | 14.7B | L | other | WestlakeNLP | 1 | 9.3k |
| EasierAI/Phi-4-14B | quantized | GGUF | — | — | mit | EasierAI | 1 | 7.8k |
| Satwik11/Microsoft-phi-4-Instruct-AutoRound-GPTQ-4bit | quantized | GPTQ | 14.7B | L | mit | Satwik11 | 2 | 6.4k |
| DevQuasar-2/microsoft.phi-4-GGUF | quantized | GGUF | — | — | — | DevQuasar-2 | 1 | 6.3k |
| Mungert/Phi-4-reasoning-GGUF | quantized | GGUF | — | — | mit | Mungert | 1 | 6.3k |
| AXCXEPT/phi-4-deepseek-R1K-RL-EZO | finetune | — | 14.7B | L | mit | AXCXEPT | 35 | 4.5k |
| gaianet/Phi-4-reasoning-GGUF | quantized | GGUF | — | — | mit | gaianet | 0 | 6k |
| pankajmathur/orca_mini_phi-4 | finetune | — | 14.7B | L | mit | pankajmathur | 8 | 5k |
| SicariusSicariiStuff/Phi-lthy4 | finetune | — | 11.9B | L | mit | SicariusSicariiStuff | 61 | 1.8k |
| Xtra-Computing/XtraGPT-14B | finetune | — | 14.7B | L | other | Xtra-Computing | 5 | 3.9k |
| nablasinc/NABLA-VL | finetune | — | 15.1B | L | apache-2.0 | nablasinc | 4 | 3.9k |
| prithivMLmods/Phi-4-o1 | finetune | — | 14.7B | L | mit | prithivMLmods | 23 | 2.7k |
| fhamborg/phi-4-4bit-gptq | quantized | -4bit | 2.85B | M | mit | fhamborg | 2 | 3.6k |
| jpacifico/Chocolatine-14B-Instruct-DPO-v1.3 | finetune | — | 14.7B | L | mit | jpacifico | 2 | 3.6k |
| tensorblock/phi-4-GGUF | quantized | GGUF | — | — | mit | tensorblock | 1 | 3.5k |
| tzervas/phi-4-bitnet-1.58b | quantized | GGUF | 3.89B | M | mit | tzervas | 2 | 3.4k |
| ykarout/phi-4-deepseek-r1-distilled-v8 | finetune | — | 14.7B | L | mit | ykarout | 0 | 3.3k |
| Date | Commit | SHA |
|---|---|---|
| 2026-07-14 | Restore <|endoftext|> (100257) as a stop token in generation_config (#67) | 2db69c1 |
| 2025-11-24 | Upload data_summary_card.md (#52) | 932b33c |
| 2025-11-24 | Update README.md (#53) | b8beb3a |
| 2025-02-24 | Remove `auto_map` so that the model is recognised as a transformers compatible model (#39) | 187ef03 |
| 2025-02-04 | Suggested tokenizer changes by Unsloth.ai (#21) | 6fbb3d3 |
| 2025-01-08 | Add Phi-4 Technical Report Link (#10) | f957856 |
| 2024-12-11 | fix(config.json): Adjusts _name_or_path value. | 381727a |
| 2024-12-11 | chore(root): Initial files upload. | 8ee52dc |
| 2024-12-11 | initial commit | 603cbf3 |
Hosted endpoints observed for phi-4, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting. Cheapest observed input price: $0.06/M.
| Provider | Model ref | Context | Quant | $/M in | $/M out | tok/s | TTFT | Tools | Status | Observed |
|---|---|---|---|---|---|---|---|---|---|---|
| DeepInfra hf, deepinfra, openrouter | microsoft/phi-4 | 16K | bfloat16 | $0.06 | $0.11 | 90.5 | 520 ms | no | live | 2026-10-25 |
| Featherless hf | microsoft/phi-4 | — | — | — | — | — | — | — | live | 2026-10-25 |
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.
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).
Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:33 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:00 UTC · Last verified 2026-10-25
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.
