The Open-Source AI Stack
RSS
All models

Models · phi

Phi-4

Open Microsoft · 2024-12-12 · MIT

14B parameters reaching benchmark scores within range of much larger models, via aggressive synthetic-data pretraining and data quality curation. The Phi line continues the "data quality over scale" thesis.

Cost

$0.13 / Mtok input
$0.50 / Mtok output

Together AI · as of 2026-05-21

via Artificial Analysis ↗

Speed

30.9 tok/sec output
542 ms TTFT

Together AI · as of 2026-05-21

via Artificial Analysis ↗

Architecture

tokens in Embedding vocab not disclosed · phi tokenizer × N layers Grouped-Query Attention RoPE context 16,384 tokens Dense MLP SwiGLU activation (standard) 14.7B active params Output projection tokens out
Schema-generated from data/models.yaml. Every label is auditable against the model's sources.

Specs

Architecture
dense
Total params
14B
Active params
14.7B
Context window
16K tokens
Attention
gqa
Position encoding
rope
Pretraining tokens
9.8T
Post-training
sft, dpo
OSI-approved
yes
Data released
no
Training code
not released

Benchmarks

Each score carries the date it was published; we never infer or interpolate missing scores.

General reasoning

MMLU 84.8 as of 2024-12-12 source ↗
MMLU-Pro 71.4 as of 2026-05-21 source ↗
GPQA-Diamond 56.1 as of 2024-12-12 source ↗

Code

HumanEval 82.6 as of 2024-12-12 source ↗
LiveCodeBench 23.1 as of 2026-05-21 source ↗

Math

MATH 80.4 as of 2024-12-12 source ↗
AIME 2024 14.3 as of 2026-05-21 source ↗
AIME 2025 18.0 as of 2026-05-21 source ↗

Recommended use cases

  • mid-tier local deployment
  • synthetic-data research reference

Available quantizations

GGUF llama.cpp's container; the common local format, k-quants from Q2 to Q8. runs on llama.cpp, Ollama
AWQ Activation-aware 4-bit weight quantization for GPU serving. runs on vLLM, SGLang
GPTQ Post-training 4-bit weight quantization for GPU serving. runs on vLLM, SGLang, Transformers
EXL2 ExLlamaV2's variable-bitrate format for consumer GPUs. runs on ExLlamaV2
MLX Apple MLX 4/8-bit layout for Apple silicon. runs on Apple MLX
FP8 8-bit float, frequently a native release on Hopper / Blackwell GPUs. runs on vLLM, SGLang, TensorRT-LLM
bitsandbytes On-the-fly NF4 / INT8 weight quantization inside Transformers. runs on Transformers

Verified via the Hugging Face model tree ↗. Community quantizations change over time; the families shown are those with published weights at audit time.

Notable innovations

  • · Synthetic-data-heavy pretraining curriculum

Known limitations

  • · Heavy synthetic-data pretraining produces benchmark scores that don't always translate to open-ended deployment. source ↗

Lineage

Continues the synthetic-data pretraining thesis from Phi-1 through Phi-3.

Sources