Gemma 3 270M
A small language model with 270 million parameters released by Google as part of the Gemma 3 series, suited to deployment on edge devices.
Timeline (last 90 days)
Setting ort.wasm.wasmPaths to point to the regular WebAssembly build before loading any model allows 4-bit Gemma 3 270M (about 323 MB) to successfully load and run on Chrome's CPU.
This bug was exposed by an app named Paste-Safe, which locally in the browser masks customer names before the user pastes text into a chat AI, and uses Gemma 3 270M for an optional second check.
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.
In an evaluation of 30 answerable questions, off-the-shelf small models under 350MB achieved faithfulness (all assertions can be found in the passages) at most 3 times.
Zero-shot Gemma 3 270M is nearly unusable on the grammar correction task, with an F0.5 of only 0.122, and 99.6% of correct sentences were altered.
All Facts (5)
Setting ort.wasm.wasmPaths to point to the regular WebAssembly build before loading any model allows 4-bit Gemma 3 270M (about 323 MB) to successfully load and run on Chrome's CPU.
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%UnverifiedIn an evaluation of 30 answerable questions, off-the-shelf small models under 350MB achieved faithfulness (all assertions can be found in the passages) at most 3 times.
50%UnverifiedZero-shot Gemma 3 270M is nearly unusable on the grammar correction task, with an F0.5 of only 0.122, and 99.6% of correct sentences were altered.
50%UnverifiedThis bug was exposed by an app named Paste-Safe, which locally in the browser masks customer names before the user pastes text into a chat AI, and uses Gemma 3 270M for an optional second check.
50%