Skip to content
TrustList

TrustList summary

Microsoft Research built Phi-4 as a small model that aims at strong reasoning without the compute demands of the largest systems. Its Hugging Face card calls it an open, dense decoder-only transformer with 14 billion parameters, a 16K token context length and an MIT licence, and it was released on 12 December 2024.

Data chosen for reasoning

The training blend, as Microsoft describes it, combines filtered public documents and code, newly created synthetic textbook-like material for maths, coding and common-sense reasoning, acquired academic books and question-and-answer sets, and chat-format supervised data covering instruction following, truthfulness and helpfulness. Multilingual text makes up about 8 percent of the whole. Training used 9.8 trillion tokens on 1920 H100 GPUs over 21 days, with a data cutoff of June 2024 for public sources. After pretraining, Microsoft applied supervised finetuning and direct preference optimization to sharpen instruction adherence and safety.

Where it is meant to be used

The card names three target conditions: memory or compute constrained environments, latency bound scenarios, and tasks needing reasoning and logic. It presents the model as a building block for generative AI features and for language model research, mainly in English. Developers are told to evaluate accuracy, safety and fairness before using it in high-risk settings.

Evaluation and safety work

Microsoft compared Phi-4 with earlier Phi models and several larger or rival models using a public simple-evals framework and its own internal tests, across MMLU, GPQA, MATH, HumanEval, MGSM, DROP and SimpleQA. The card flags factual recall (SimpleQA) as an area where it trails. Safety post-training included red-teaming against jailbreaks and multi-turn attacks. Prompts work best in chat format with system, user and assistant turns, and the card shows loading it through the transformers pipeline. It also warns that non-English output will be weaker.

Summary written by TrustList from the business's own published information, Nov 2026.

Benchmarks & AI stats

BenchmarkOfficialCommunity avg
No benchmark scores yet. Be the first to add one.

“Official” values are editor-approved and feed the ranking. “Community avg” is the mean of member submissions (shown for transparency; it never affects the ranking until an editor approves a value).

Sign in to add a benchmark score for this model.

Phi-4 launches

Request a demo or quote from Phi-4

Protected by reCAPTCHA - Google Privacy Policy and Terms apply.

By sending, you agree we may share your request and contact details with the provider once you confirm your email.

Reviews

Write the first review of Phi-4

Used it? Your experience helps other buyers decide.

Write a review

Questions & answers

No questions yet. Be the first to ask about Phi-4.

Phi-4 alternatives

Similar listings buyers compare against this one.

Google's open multimodal embedding model that maps text, images, audio and video into one vector space, under Apache 2.0

Microsoft AI's real-time speech-to-text model for voice agents, transcribing 60 languages with continuous language detection

Cantina Security's first open-weight security model, a post-trained GLM-5.3-Flash for vulnerability research and exploit checks

Open-source 2B decision model from AWS's Strands Agents project that picks options and returns scores with a reliability figure

Perplexity's open-weight decision model that picks among given options for routing and checks in AI systems

Open-weight translation model from Bilibili's Index team, covering 150 languages and following glossary and style instructions

Aleph Alpha's open-weight English-German mixture-of-experts model for self-hosting, released under Apache 2.0 in October 2026

Microsoft AI's text-to-speech model for keeping one voice across 23 languages, with voice cloning from short audio