Phi 4 Multimodal Instruct
Microsoft · released Feb 24, 2025 · microsoft/Phi-4-multimodal-instruct
Speech to textTranscribes a recording into words
- Type
- Open weightsMIT License
- Languages
- 23
- Size
- 5.6B
about 98K words of context
Our take
Written Sep 4, 2026Phi 4 Multimodal Instruct is a small downloadable speech-to-text model from Microsoft with a permissive MIT licence. It transcribes faster than most models we track and keeps fewer words wrong than the median across every measured audio condition, though it is not the single most accurate on any of them.
Pick this for fast local transcription where you need open weights with no licence strings attached — it processes an hour of audio in 22 seconds on the benchmark hardware. Use it for clean read-aloud work, podcasts, meetings or accented speech, where it beats the median model on every condition. Skip it if you need a hosted provider, speaker diarisation, timestamps, or the absolute lowest error rate on any single audio type.
The case for it
- Processes audio at 163 times real time — faster than most models on the benchmark harness.
- Beats the median error rate on every measured condition: clean read-aloud, podcasts and video, accented speech, and meetings.
- MIT licence with no attribution or copyleft requirements.
- 23 languages supported, unusually broad for a 5.6-billion-parameter model.
The case against it
- No commercial hosting options we list — you must run it yourself or self-host.
- Not the most accurate on any single condition; the field leader reaches 0.9% on clean read-aloud versus its 1.4%.
- Every accuracy figure is English-only; nothing we hold measures the other 22 languages.
How good is it?
TranscriptionTurning speech into text3.5 of 5Open ASR WER · 33rd of 76
95%
Misses roughly one word in 20, averaged over nine English test sets.
163×60th of 74
an hour of audio in 22 seconds, on the board's own hardware. Your machine will differ.
23
Listed on the model card. The accuracy above is English only.
Percentage of words wrong on each set, lower better. Bars are scaled to this model's own worst case; the placing beneath each rate is against every model measured on that set.
Which languages ↓ ↑
Arabic · Chinese · Czech · Danish · Dutch · English · Finnish · French · German · Hebrew · Hungarian · Italian · Japanese · Korean · Norwegian · Polish · Portuguese · Russian · Spanish · Swedish · Thai · Turkish · Ukrainian
The figures above come from the Open ASR Leaderboard, an independent public test that runs every model on the same recordings. It is the only measurement of transcription quality we know of, so there are no other scores to show.
Each of these is the same transcription job on a different kind of recording, so together they say where it holds up and where it slips — not how closely it follows an instruction.
Every published score for this model9 scoresEvery figure we hold, from 9 boards, with who ran it and a link to the source — including the boards no rating above is built on.
Can you run it yourself?
Comfortable fit
GeForce RTX 4090 · 24 GB
Room to spare. 17.7 GB spare means a 10% error in the size would not change the answer.
GeForce RTX 5090 · 32 GB
Room to spare. 25.7 GB spare means a 10% error in the size would not change the answer.
Apple M2 (8-core GPU, 8GB unified) · 8 GB
Room to spare. 0.9 GB spare means a 10% error in the size would not change the answer.
These cards answer whether Phi 4 Multimodal Instruct loads, not how fast it transcribes. Throughput figures for a transcription model come from its text decoder, so treat this as a fit answer rather than a speed one.
Memory use by level
Against a 24 GB card.
What is quantisation? →This model on every device we track71 devicesThe Q4 build most people download, on each device: what the weights come to, how much context the memory leaves, and whether it runs. Smallest device that runs it first. This is a fit answer: whether it loads, not how fast it transcribes.
When we formed this view
Recent changes
What moved
first indexed by our pipelineEach date is the day we first saw the change, or the day the maker announced it.
What we do not know about this model yet
- We hold no measured file for it, so all 3 sizes on this page are calculated from the parameter count.
- Nothing we hold says whether an endpoint streams, so we do not show it either way.
Licence and identifiers
What the licence allowsMIT License, what it allows commercially, and the identifiers you need to pull this model — its Hugging Face repo, our slug and a machine-readable card.
Licence
MIT License
Fully permissive: do anything with attribution. No patent grant, unlike Apache-2.0.
Identifiers
- Hugging Face
- microsoft/Phi-4-multimodal-instruct
- Architecture
- Dense
- Takes in, gives back
- Audio in, text out
- Catalogue slug
- microsoft-phi-4-multimodal-instruct