Phi-4

AI model Released Dec 12, 2024

Context window: 16,384 tokens^*
Total parameters: 14.66 B^*
Open weights: Yes^*

Capability & preference

General capability · ECI

A cross-benchmark capability estimate. Index points, not a percentage; intervals show uncertainty.

Phi-4 129.2^* Phi-4 · 95% interval 126.0–132.1

Run it locally

Inference · memory estimate

Phi-4

14.66 B^* total parameters · phi3^* · BF16^*

Published checkpoint tensor payload: 27.3^* GiB. This includes stored copies and buffers; it is separate from the runtime estimates below.

Weight precision Total context Check a GPU Check a Local AI system

Estimated accelerator memory9.3^* GiB

Weights: 7.5^* GiB
KV cache: 0.8^* GiB
Runtime allowance: 1.0^* GiB

One user, full model in accelerator memory.

GPUs with memory headroom

Local AI systems with memory headroom

NVIDIA Jetson AGX Orin Developer Kit (64 GB) Planning budget: 48.0 GiB
AMD Ryzen AI Halo (Ryzen AI Max+ 395, 128 GB) Planning budget: 96.0 GiB
ASUS Ascent GX10 (128 GB) Planning budget: 96.0 GiB
Apple Mac Studio (M5 Max 18-core CPU, 40-core GPU, 128 GB) Planning budget: 96.0 GiB
Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB) Planning budget: 96.0 GiB
Dell Pro Precision with GB10 FCM1253 (128 GB, 2 TB) Planning budget: 96.0 GiB

Shared-memory systems use a planning budget of 75% of installed memory, leaving 25% for the OS and other allocations. This is an assumption, not a measured GPU allocation. Firmware and driver limits may be lower and may require configuration. Memory matches do not verify a runtime, quantized model artifact or inference speed.

Includes FP16 KV cache, prompt plus generated tokens, quantization metadata and a runtime allowance. GPU matches reserve 10% of VRAM. Memory fit is not a speed or software-compatibility guarantee; quantized weights and runtime support must be available.

Capabilities

Context window: 16,384 tokens^*
Open weights: Yes^*
Model license: mit^*
Total parameters: 14.66 B^*
Checkpoint tensor payload: 27.31 GiB^*

Input modalities

text^*

Output modalities

text^*

Processing & platform

Architecture: phi3^*

System configuration

Detailed specifications

Specifications 1 · 4 fields

Access model: Open weights (unrestricted)^*
Open weights: Yes^*
Organization: Microsoft Research^*
Organization country: United States of America^*

Specifications 2 · 1 fields

Checkpoint memory architecture

Unique parameters: 14,659,507,200^*
Stored tensor elements: 14,659,507,200^*
Checkpoint tensor bytes: 29,319,014,400^*
Checkpoint data type: BF16^*
Architecture: phi3^*
Context limit: 16,384^*
Transformer layers: 40^*
Key/value heads: 10^*
Attention head dimension: 128^*
Cache layout: full-attention^*

Specifications 3 · 22 fields

Architecture: phi3^*
Attention bias: No^*
Attention dropout: 0.0^*
Activation function: silu^*
Hidden dimension: 5,120^*
Feed-forward dimension: 17,920^*
Attention heads: 40^*
Transformer layers: 40^*
Key/value heads: 10^*
RMS normalization epsilon: 1.0e-5^*
RoPE base: 250,000^*
Tied word embeddings: No^*
Checkpoint data type: bfloat16^*
Vocabulary size: 100,352^*
Context window: 16,384^* tokens
Checkpoint input modalities: text^*
Checkpoint licence: mit^*
Checkpoint output modalities: text^*
Total parameters: 14.6595072^* B
Checkpoint tensor bytes: 29,319,014,400^* bytes
Saved tensor elements: 14,659,507,200^* elements
Checkpoint tensor payload: 27.30546^* GiB

* Provisional data