# Low-latency Inference

Arch-Router is a preference-aligned routing framework model with 1.5B parameters, used to map queries to domain-operation preferences for model routing decisions.

DMind-1 is a Web3 expert model built upon Qwen3-32B, optimized for the Web3 ecosystem through supervised instruction fine-tuning and human feedback reinforcement learning, achieving significant improvements in task accuracy, content safety, and expert-level interaction alignment.

TreeHop is a lightweight embedding-level framework designed for efficient query embedding generation and filtering in multi-hop QA, significantly reducing computational overhead.

Distil-Whisper is a distilled version of the Whisper model, achieving efficient speech recognition through large-scale pseudo-labeling technology

NVIDIA NeMo Canary Flash is a multilingual multitask speech model supporting automatic speech recognition and translation tasks in English, German, French, and Spanish.

The FP8 quantized version of Qwen2.5-VL-3B-Instruct, supporting visual-text input and text output, and optimizing inference efficiency.

Mistral Small 3 (version 2501) is a 24B-parameter instruction-tuned large language model that sets a new benchmark in the sub-70B parameter category, featuring exceptional knowledge density and multilingual support capabilities.

A lightweight Counter-Strike 2 player detection model based on YOLOv11, suitable for real-time object detection scenarios

This is a Transformer-based Reranker model primarily used for information retrieval and search result optimization tasks.

This is a Japanese speech-to-any-target-language text translation pipeline based on a cascaded approach, consisting of automatic speech recognition (ASR) and text translation components.

Kotoba-Whisper-v2.1 is a Japanese automatic speech recognition (ASR) model based on Whisper, integrating an additional post-processing stack that automatically adds punctuation marks.

This is a Transformers-based Text-to-Speech (TTS) model capable of converting input text into natural speech output.

MobileCLIP-S1 is an efficient image-text model that achieves fast zero-shot image classification through multimodal reinforcement training.

This model is a fine-tuned audio classification model based on asapp/sew-mid-100k on the alexandreacff/kaggle-fake-detection dataset, designed for fake audio detection.

Counter-Strike 2 (CS2) player object detection model based on YOLOv9 architecture, capable of recognizing player characters in the game

This is a keypoint detection model based on transformers, used to identify keypoint positions in images

Mixtral-8x22B-Instruct-v0.1 is a large language model fine-tuned for instructions based on Mixtral-8x22B-v0.1, supporting multiple languages and function calling capabilities.

Japanese-optimized version based on OpenAI Whisper large-v3, supporting multilingual speech recognition

Whisper large-v2 is a large-scale automatic speech recognition (ASR) model developed by OpenAI, supporting multilingual speech-to-text tasks.

This is an audio conversion model based on RVC (Retrieval-based Voice Conversion) technology, capable of transforming input audio into speech with a specific style.

This is an RVC (Retrieval-Based Voice Conversion) model designed for audio-to-audio conversion tasks.

Sonic48k is an audio-to-audio model based on RVC (Retrieval-based Voice Conversion) technology, primarily used for voice conversion tasks.

This is an audio conversion model based on RVC (Retrieval-Based Voice Conversion) technology, capable of converting input audio into a specific character's voice.

This is a voice conversion model based on RVC (Retrieval-based Voice Conversion) technology, capable of converting input audio into a specific style of speech.

This is an RVC (Retrieval-Based Voice Conversion) model designed for audio-to-audio tasks, particularly suited for converting speech into the Lego City narrator style.

This is an audio conversion model based on RVC (Retrieval-based Voice Conversion) technology, capable of converting input audio into the voice of Jesse Pinkman.

This is a voice conversion model based on RVC (Retrieval-Based Voice Conversion) technology, capable of converting input audio into a specific style of voice output.

A Japanese automatic speech recognition (ASR) model based on Whisper Large V2, optimized with CTranslate2 for efficient inference.

This model is a DistilBERT model fine-tuned on the SQuAD dataset for extractive question answering tasks, with high exact match rate and F1 score.

LeViT-256 is an efficient vision model based on Transformer architecture, designed for fast inference and pretrained on the ImageNet-1k dataset.

MobileViT is a lightweight, low-latency vision Transformer model that combines the strengths of CNNs and Transformers, making it suitable for mobile devices.

MobileViT is a lightweight, low-latency vision Transformer model that combines the advantages of CNNs and Transformers, suitable for mobile devices.

A lightweight sentence similarity model optimized for financial domain question matching, compressing 12-layer BERT to 4 layers through distillation technology, significantly improving inference efficiency

A cross-encoder model trained on the MS Marco passage ranking task, suitable for query-passage relevance scoring in information retrieval scenarios.

A versatile large language model capable of handling various natural language processing tasks (inferred information)

DistilBART is a distilled version of the BART model, specifically optimized for summarization tasks, significantly reducing model parameters and inference time while maintaining high performance.

A Korean extractive Q&A model based on KLUE-BERT-base, specifically designed for Korean machine reading comprehension tasks.