Supported Models

All supported models are hosted under the inference4j HuggingFace organization and are automatically downloaded and cached on first use.

NLP

Capability	Wrapper	Default Model ID	Size	API
Text Classification	DistilBertTextClassifier	inference4j/distilbert-base-uncased-finetuned-sst-2-english	~260 MB	TextClassifier
Text Embeddings	SentenceTransformerEmbedder	inference4j/all-MiniLM-L6-v2	~90 MB	TextEmbedder
Search Reranking	MiniLMSearchReranker	inference4j/ms-marco-MiniLM-L-6-v2	~90 MB	SearchReranker
Text Generation	OnnxTextGenerator.gpt2()	inference4j/gpt2	~500 MB	TextGenerator
Text Generation	OnnxTextGenerator.smolLM2()	inference4j/smollm2-360m-instruct	~700 MB	TextGenerator
Text Generation	OnnxTextGenerator.smolLM2_1_7B()	inference4j/smollm2-1.7b-instruct	~3.4 GB	TextGenerator
Text Generation	OnnxTextGenerator.tinyLlama()	inference4j/tinyllama-1.1b-chat	~2.2 GB	TextGenerator
Text Generation	OnnxTextGenerator.qwen2()	inference4j/qwen2.5-1.5b-instruct	~3 GB	TextGenerator
Text Generation	OnnxTextGenerator.gemma2()	Gated — requires manual download	~5 GB	TextGenerator
Summarization / Translation / Grammar	FlanT5TextGenerator.flanT5Small()	inference4j/flan-t5-small	~300 MB	TextGenerator, Summarizer, Translator, GrammarCorrector
Summarization / Translation / Grammar	FlanT5TextGenerator.flanT5Base()	inference4j/flan-t5-base	~900 MB	TextGenerator, Summarizer, Translator, GrammarCorrector
Summarization / Translation / Grammar	FlanT5TextGenerator.flanT5Large()	inference4j/flan-t5-large	~3 GB	TextGenerator, Summarizer, Translator, GrammarCorrector
Text-to-SQL	T5SqlGenerator.t5SmallAwesome()	inference4j/t5-small-awesome-text-to-sql	~240 MB	TextGenerator, SqlGenerator
Text-to-SQL	T5SqlGenerator.t5LargeSpider()	inference4j/T5-LM-Large-text2sql-spider	~4.6 GB	TextGenerator, SqlGenerator
Summarization	BartSummarizer.distilBartCnn()	inference4j/distilbart-cnn-12-6	~1.2 GB	TextGenerator, Summarizer
Summarization	BartSummarizer.bartLargeCnn()	inference4j/bart-large-cnn	~1.6 GB	TextGenerator, Summarizer
Translation	MarianTranslator.builder()	User-specified (inference4j/opus-mt-*)	varies	TextGenerator, Translator
Grammar Correction	CoeditGrammarCorrector.coeditBase()	inference4j/coedit-base	~900 MB	TextGenerator, GrammarCorrector
Grammar Correction	CoeditGrammarCorrector.coeditLarge()	inference4j/coedit-large	~3 GB	TextGenerator, GrammarCorrector

Vision

Capability	Wrapper	Default Model ID	Size	API
Image Classification	ResNetClassifier	inference4j/resnet50-v1-7	~100 MB	ImageClassifier
Image Classification	EfficientNetClassifier	inference4j/efficientnet-lite4	~50 MB	ImageClassifier
Object Detection	YoloV8Detector	inference4j/yolov8n	~25 MB	ObjectDetector
Object Detection	Yolo26Detector	inference4j/yolo26n	~25 MB	ObjectDetector
Text Detection	CraftTextDetector	inference4j/craft-mlt-25k	~80 MB	TextDetector

Multimodal

Capability	Wrapper	Default Model ID	Size	API
Zero-Shot Classification	ClipClassifier	inference4j/clip-vit-base-patch32	~595 MB	ZeroShotClassifier
Image Embeddings	ClipImageEncoder	inference4j/clip-vit-base-patch32	~340 MB	ImageEmbedder
Text Embeddings (CLIP)	ClipTextEncoder	inference4j/clip-vit-base-patch32	~255 MB	TextEmbedder

Audio

Capability	Wrapper	Default Model ID	Size	API
Speech-to-Text	Wav2Vec2Recognizer	inference4j/wav2vec2-base-960h	~370 MB	SpeechRecognizer
Voice Activity Detection	SileroVadDetector	inference4j/silero-vad	~2 MB	VoiceActivityDetector

Generative AI

Generative models use a separate module (inference4j-genai) and a different builder pattern. See Generative AI for details.

Capability	Wrapper	Model ID	Size	License
Text Generation	TextGenerator	inference4j/phi-3-mini-4k-instruct	~2.7 GB	MIT
Text Generation	TextGenerator	inference4j/deepseek-r1-distill-qwen-1.5b	~1 GB	MIT
Speech-to-Text / Translation (WIP)	WhisperSpeechModel	inference4j/whisper-small-genai	~500 MB	MIT
Vision-Language	VisionLanguageModel	inference4j/phi-3.5-vision-instruct	~3.3 GB	MIT

Model reference

A comprehensive view of all supported models, organized by architecture:

Encoder-only (single-pass)

Model	Tokenizer	Wrapper	Use Cases
DistilBERT SST-2	WordPiece	DistilBertTextClassifier	Sentiment analysis, text classification
all-MiniLM-L6-v2	WordPiece	SentenceTransformerEmbedder	Semantic search, embeddings
MiniLM-L-6 MS MARCO	WordPiece	MiniLMSearchReranker	Search reranking

Decoder-only (autoregressive)

Model	Tokenizer	Wrapper	Use Cases
GPT-2	BPE	OnnxTextGenerator	Text completion
SmolLM2-360M	BPE	OnnxTextGenerator	Chat, instruction following
TinyLlama-1.1B	SentencePiece BPE	OnnxTextGenerator	Chat, instruction following
Qwen2.5-1.5B	BPE	OnnxTextGenerator	Chat, instruction following
Gemma 2-2B	SentencePiece BPE	OnnxTextGenerator	Chat, instruction following

Encoder-decoder (autoregressive)

Model	Tokenizer	Wrapper	Use Cases
Flan-T5 (Small / Base / Large)	SentencePiece Unigram	FlanT5TextGenerator	Summarization, translation, grammar
T5-small-awesome-text-to-sql	SentencePiece Unigram	T5SqlGenerator	Text-to-SQL (lightweight)
T5-LM-Large-text2sql-spider	SentencePiece Unigram	T5SqlGenerator	Text-to-SQL (high accuracy)
DistilBART CNN 12-6	BPE	BartSummarizer	Summarization
BART Large CNN	BPE	BartSummarizer	Summarization
MarianMT (opus-mt-*)	SentencePiece BPE	MarianTranslator	Translation (fixed language pair)
CoEdIT (Base / Large)	SentencePiece Unigram	CoeditGrammarCorrector	Grammar correction

Vision

Model	Tokenizer	Wrapper	Use Cases
ResNet-50	N/A	ResNetClassifier	Image classification
EfficientNet-Lite4	N/A	EfficientNetClassifier	Image classification
YOLOv8n	N/A	YoloV8Detector	Object detection
YOLO26n	N/A	Yolo26Detector	Object detection
CRAFT	N/A	CraftTextDetector	Text detection in images

Multimodal

Model	Tokenizer	Wrapper	Use Cases
CLIP ViT-B/32	BPE	ClipClassifier	Zero-shot image classification

Audio

Model	Tokenizer	Wrapper	Use Cases
Wav2Vec2	CTC	Wav2Vec2Recognizer	Speech-to-text
Silero VAD	N/A	SileroVadDetector	Voice activity detection

Model compatibility

YOLOv8 / YOLO11

YoloV8Detector is compatible with both YOLOv8 and YOLO11 models — they share the same output layout ([1, 4+C, N]). It is not compatible with YOLOv5 (different layout with objectness column) or YOLO26 (NMS-free architecture).

EfficientNet variants

EfficientNetClassifier is tested against EfficientNet-Lite4 (TensorFlow origin, softmax built-in). For PyTorch-exported EfficientNet models that output raw logits, override with .outputOperator(OutputOperator.softmax()).

Custom models

Any ONNX-exported model works with the appropriate wrapper, as long as it follows the expected input/output layout. See the Custom Models guide for details.

Cache

Models are cached in ~/.cache/inference4j/ by default. Customize with:

• System property: -Dinference4j.cache.dir=/path/to/cache
• Environment variable: INFERENCE4J_CACHE_DIR=/path/to/cache

See Configuration for all options.

Planned models

Domain	Model	Status
Text	TrOCR (text recognition)	Planned — enabled by encoder-decoder infrastructure

See the Roadmap for details.