By ByteBot
ONNX just made its most meaningful release in years. Version 1.22.0, published June 30, introduces Opset 27 with two new operators — LinearAttention and CausalConvWithState — that let modern hybrid LLMs export cleanly to ONNX for the first time. Tack on browser-based model validation via WebAssembly and cryptographic supply-chain provenance baked into every release artifact, and you have a release worth actually reading the changelog on.
The Attention Gap, Finally Closed
Here is the problem ONNX has quietly had for the last two years: it could not represent transformer attention semantics cleanly for modern architectures. The original Attention operator predates grouped-query attention, linear attention, and the hybrid recurrent-softmax designs that power efficient LLMs in 2026. When you exported a model using any of those patterns, you either reached for custom ops — which are tied to specific ONNX Runtime versions and break portability — or you flattened the architecture to something ONNX could express, which meant losing the very efficiency gains that made the model worth deploying in the first place.
Opset 27 fixes this with two new operators. LinearAttention-27 implements linear (O(n)) recurrent attention. CausalConvWithState-27 handles causal convolution with persistent state, used in hybrid architectures. Together, they cover the building blocks of the modern efficient-LLM generation.
The practical proof is already on Hugging Face: the Qwen3.5-0.8B ONNX community model is live. Qwen3.5 0.8B uses Gated DeltaNet — a 3:1 ratio of linear-to-softmax attention layers — which is exactly why it runs at 403.5 tokens/second at $0.02 per million tokens on DeepInfra. Before v1.22, cleanly exporting that architecture required relying on runtime-specific contrib ops with no spec guarantee. Now the semantics are in the standard.
There is also a correctness fix bundled here: the function definitions for Attention-23 and Attention-24 had bugs under grouped-query attention mode. Most modern LLMs use GQA. If you have been exporting GQA models and getting subtly wrong outputs in ONNX, this is why, and v1.22 fixes it. Upgrade for correctness alone if nothing else.
Browser Validation: Useful in One Direction
ONNX v1.22.0 adds WebAssembly support via Pyodide, which means you can validate and inspect ONNX models in a browser with no server and no local Python install. Pyodide compiles CPython to WebAssembly, so the full ONNX checker runs client-side.
The use cases are real but specific: PR review workflows where a reviewer needs to validate a model artifact before merge; compliance audits in regulated industries where the person reviewing a model is not an engineer and does not have a Python environment; and ML courses that want live model validation without infrastructure dependencies.
A word on what this is not: it is not a replacement for ONNX Runtime Web for actual inference. Running a CPython interpreter inside WASM carries interpreter overhead on top of WASM’s own constraints — it is slow. For inference in the browser you still want onnxruntime-web directly, ideally with WebGPU for anything non-trivial. The Pyodide integration handles validation and inspection, and it does that job well.
Supply-Chain Security: ONNX Is First
Every ONNX v1.22.0 release artifact now ships with SLSA Level 2 provenance attestations and a full embedded SBOM. This is a bigger deal than it sounds, and ONNX is the first ML exchange format to do it.
SLSA Level 2 means every artifact is cryptographically traceable back to a version-controlled source, built on a tracked build service with signed metadata. The SBOM documents exactly what went into the release. When a supply-chain incident hits the ML ecosystem — and in 2026, HuggingFace pickle exploits are not a theoretical concern, they are documented — teams with provenance attestations know within hours what they are exposed to. Teams without them spend weeks on manual audits.
For anyone deploying in a regulated environment — finance, healthcare, defense — this changes ONNX from a convenient interoperability layer into an auditable deployment artifact. TensorRT, ExecuTorch, and LiteRT do not offer this. If your compliance team has been pushing back on ML dependencies, v1.22 gives you something concrete to point to.
When to Upgrade
If you are deploying any model with linear attention, grouped-query attention, or hybrid recurrent-softmax architecture, upgrade to Opset 27 now — both for the new operators and for the GQA correctness fixes. If you are in a regulated environment with supply-chain requirements, this release gives you the provenance trail you need. If you are doing browser-based ML tooling, the Pyodide validation support is a clean addition.
The v1.22.0 release notes and full changelog are on the ONNX GitHub releases page. The official announcement is on the LFAI & Data blog. If you are also on the training side, this pairs well with what landed in PyTorch 2.13’s FlexAttention — train with flexible attention patterns in PyTorch, then export them cleanly in ONNX.
