Skip to content
English

Switch theme

// Phi · Speech-to-text

Phi-4-multimodal-instruct: VRAM & local deployment

5.57B parameters. About 4.2 GB at 4-bit and 12.8 GB at BF16.

released 2025-02-24licence: Commercial OKmemory figures are estimates

01Specs

PublisherMicrosoft · microsoft
Family / seriesPhi · Phi-4-multimodal
TypeSpeech-to-text (automatic-speech-recognition)
Parameters5.57B Source: counted from official safetensors weights
Released weightsBF16 · 11.2 GBdownload
LicenceMIT · Commercial OKPermissive licence (Apache-2.0 / MIT / CC-BY / OpenMDW): commercial use allowed.
Released2025-02-24 (Hugging Face repo created)

02How much VRAM it needs (estimate)

Speech models are small; the weight size is close to the total memory needed.

PrecisionWeightsWeights + overheadSmallest GPUMac
BF16 / FP16 (original)11.2 GB 12.8 GB 16GB GPUMac 24GB
FP8 / INT8 / Q8_05.9 GB 7.0 GB 8GB GPUMac 16GB
4-bit (Q4_K_M / AWQ / GPTQ-Int4)3.4 GB 4.2 GB 8GB GPUMac 16GB

Bits per weight: BF16 16, FP8 8.5, 4-bit 4.85. “Fits” = total ≤ 90% of memory; Mac GPU can use ~2/3 of unified memory (≤36GB) or ~3/4 (larger). For MoE, all experts must be in memory, so size by total parameters. Method →

03How to run it

Only runtimes the publisher or the runtime’s official library names for this model.

Which inference engine to choose →

04Official repositories & quantised builds

2 official repo(s) from microsoft.

05Other Phi models

All 18 →
ModelParams4-bit totalReleased
Phi-4-reasoning-vision-15B15.1B12.3 GB2026-01-23
Phi-mini-MoE-instruct7.6B / 2.4B6.6 GB2025-06-23
Phi-tiny-MoE-instruct3.8B / 1.1B3.6 GB2025-06-23
Phi-4-mini-flash-reasoning3.85B4.4 GB2025-06-19
Phi-4-mini-reasoning3.84B4.1 GB2025-04-29
Phi-4-reasoning-plus14.7B12.0 GB2025-04-17
Phi-4-reasoning14.7B12.0 GB2025-04-09
Phi-4-mini-instruct3.84B4.1 GB2025-02-19

06Sources

Data checked 2025-12-10. Parameters, context and licence are copied from the official repository; memory figures are our estimates.

Data updated: Self-host open models(UTC+8)