Phi-4-reasoning

Checkpointmicrosoft/Phi-4-reasoningModel card indexedModel card

Support summary

Phi-4-reasoning is a checkpoint published by Microsoft, built on the Phi3ForCausalLM architecture. vLLM 0.28.0, the newest indexed vLLM release, documents support. TensorRT-LLM lists its architecture, but not this exact checkpoint.

vLLM documents support in 10 of 20 indexed releases, from 0.22.0 to 0.28.0.

Documented hardware
CPU
Documented quantization
BF16
Configured context
32,768 tokens
License
mit

Hardware, quantization and features count only statements about this checkpoint; engine-wide capabilities are listed with each release.

Hugging Face model card
Publisher
Microsoft
Task
text generation
License
mit
Library
transformers
Architecture
Phi3ForCausalLM
Weight tensor elements
14.7B
Tensor types
BF16
Configured context
32,768 tokens
Languages
en
Base models
microsoft/phi-4
Weight access
Public
Repository created
2025-04-09
Source 1: microsoft/Phi-4-reasoning · publisher model cardSource 2: microsoft/Phi-4-reasoning · Hugging Face metadataSource 3: microsoft/Phi-4-reasoning · publisher configurationReviewed
Model notes and sources

Engine support

3 engines

Engine support for Phi-4-reasoning
EngineReleaseCheckpoint supportEvidence scopeLast known supported
vLLMVersion not pinnedRecipe / cookbookDocumented recipe

Official recipe

Source: vllm · models/microsoft/Phi-4-mini-instruct.yamlReviewed
View 0.28.0
Features, dependencies and evidence for vLLM 0.28.0
SGLang0.5.19Newest indexedNo matching evidence curatedNo supported release indexed
Features, dependencies and evidence for SGLang 0.5.19
TensorRT-LLM1.2.1Newest indexedArchitecture listed; exact checkpoint unknownNo supported release indexed
Features, dependencies and evidence for TensorRT-LLM 1.2.1

Recipes and cookbooks