ONNX
ONNX (Open Neural Network Exchange) is an open standard jointly initiated by Microsoft and Facebook that serves as a common intermediate format for representing machine learning models. It defines a unified operator set and data type specification, enabling models to be converted between different training frameworks such as PyTorch and TensorFlow, and deployed across diverse hardware platforms via inference engines such as ONNX Runtime, thereby addressing interoperability challenges in the deep learning ecosystem.
Core Facts
ONNX Runtime is an open-source machine learning inference and training acceleration engine primarily developed under Microsoft's leadership.
ONNX was originally co-initiated by Microsoft and Facebook (now Meta) in 2017.
The ONNX (Open Neural Network Exchange) format allows PyTorch/TensorFlow models to be exported as a cross-platform intermediate representation format.
Timeline (last 90 days)
ONNX (Open Neural Network Exchange) is an open model exchange format jointly launched by companies such as Microsoft and Meta.
ONNX functions like a universal translator for AI models: no matter whether a model was originally trained with PyTorch, TensorFlow, or another framework, it can be exported to the ONNX format and then loaded and executed in a runtime environment that supports ONNX.
Candidate small models include SmolLM2 (135M and 360M), Gemma 3 270M, LFM2.5-350M, Qwen2.5-0.5B, and Qwen3-0.6B, all of which have ONNX browser builds.
The differences between the ONNX path and the TFLite path are not only in quantization but also in operator sets, post-processing implementations, and inference scheduling, so the score distributions of the same model on the two platforms may differ; when porting across platforms, thresholds should be recalibrated rather than directly reusing desktop-side parameters.
The iOS export uses Core ML 8-bit weights, with a model size of 26MB and 99.96% agreement with fp32 labels, and in the keyboard it actually uses the CPU compute unit, taking about 4ms per sentence in a release build; the Android version uses ONNX opset 17 with int8 dynamic quantization, 27MB, 99.1% agreement, and likewise about 4ms per sentence on CPU.
The ONNX file provides a q4f16 version and a portable q4 build; fp16 and fp32 versions are omitted to fit within Hugging Face Space's 1GB limit.
ONNX Runtime is an open-source inference engine led by Microsoft that supports exporting models from frameworks such as PyTorch and TensorFlow to the unified ONNX format for cross-platform execution, with built-in hardware backends including CPU, CUDA, and TensorRT.
Talus is exported in ONNX format with weights stored in fp16 and converted back to fp32 on load, runs on WebGPU via ONNX Runtime Web, generates one map in approximately 3 seconds on an RTX 5060, and provides a CPU fallback.
A common engineering practice is to train models in Python, export them to ONNX or TensorFlow.js format, and then have the JS frontend load and run them.
Core ML allows developers to convert models trained in formats such as ONNX and PyTorch into a format that can perform efficient inference on Apple devices without requiring network requests.
23 more timeline events
All Facts (20)
ONNX Runtime is an open-source machine learning inference and training acceleration engine primarily developed under Microsoft's leadership.
75%VerifiedONNX was originally co-initiated by Microsoft and Facebook (now Meta) in 2017.
70%VerifiedThe ONNX (Open Neural Network Exchange) format allows PyTorch/TensorFlow models to be exported as a cross-platform intermediate representation format.
65%UnverifiedMoS-TTS-Nano使用ONNX（Open Neural Network Exchange）格式进行CPU上的高效推理
95%UnverifiedTensorFlow导出ONNX需要使用tf2onnx库
95%UnverifiedTransformers框架同时支持PyTorch和TensorFlow后端，并通过ONNX等格式支持更广泛的部署场景
90%UnverifiedONNX模型导出后，只要保持相同的input_names和output_names，推理代码就是完全通用的，无论模型来自PyTorch还是TensorFlow
90%UnverifiedTensorFlow导出ONNX时不需要提供dummy数据，因为tf.function通过AutoGraph将Python代码编译为静态图，input_signature提供的类型和形状信息足以完整描述计算图结构
90%Unverifiedonnxruntime-genai包可以用于ONNX格式模型的生成式AI推理
85%Unverified2019年，ONNX被移交给Linux基金会旗下的LF AI & Data基金会管理
85%UnverifiedONNX（Open Neural Network Exchange）诞生于2017年，最初由Facebook（Meta）和Microsoft联合开发
85%Unverified许多HuggingFace模型仓库已经提供了ONNX版本，特别是针对CPU和移动端优化的量化模型
80%UnverifiedONNX已获得Amazon、Intel、NVIDIA、Qualcomm等数十家主流硬件和软件厂商的支持
80%Partially VerifiedONNX由微软和Facebook于2017年联合发起，目前已成为AI模型互操作的事实标准
75%UnverifiedTensorRT是NVIDIA推出的高性能深度学习推理优化器与运行时，自2015年发布
70%UnverifiedONNX (Open Neural Network Exchange) is an open model exchange format jointly launched by companies such as Microsoft and Meta.
50%UnverifiedONNX functions like a universal translator for AI models: no matter whether a model was originally trained with PyTorch, TensorFlow, or another framework, it can be exported to the ONNX format and then loaded and executed in a runtime environment that supports ONNX.
50%UnverifiedCandidate small models include SmolLM2 (135M and 360M), Gemma 3 270M, LFM2.5-350M, Qwen2.5-0.5B, and Qwen3-0.6B, all of which have ONNX browser builds.
50%UnverifiedThe differences between the ONNX path and the TFLite path are not only in quantization but also in operator sets, post-processing implementations, and inference scheduling, so the score distributions of the same model on the two platforms may differ; when porting across platforms, thresholds should be recalibrated rather than directly reusing desktop-side parameters.
50%UnverifiedThe iOS export uses Core ML 8-bit weights, with a model size of 26MB and 99.96% agreement with fp32 labels, and in the keyboard it actually uses the CPU compute unit, taking about 4ms per sentence in a release build; the Android version uses ONNX opset 17 with int8 dynamic quantization, 27MB, 99.1% agreement, and likewise about 4ms per sentence on CPU.
50%