webbrain-one/d1-browser-decision-fp16

d1 Browser Decision FP16

The FP16-weight version of webbrain-one/d1-browser-decision-fp32 (revision 71866f0ccab0b9fa92c6df64d1c34eacf2a7b2b8), an experimental, audio-free derivative of LiquidAI/d1-omni-600M. It scores named options for choice, noul and ordinal score questions using text/JSON state and optional screenshots. It does not generate text or tokens.

This is the decision model the WebBrain extension downloads for its optional on-device decision provider.

What changed from the FP32 release

• Every FP32 weight tensor with at least 4,096 elements in the three ONNX graphs is stored as FP16 and cast back to FP32 where it is used. The token-embedding table is gathered in FP16 and only the gathered rows are cast.
• Activations, attention, softmax and accumulation stay FP32. Tokenizer, calibration temperatures, preprocessing and the answer semantics are unchanged.
• No training, merge, selection or PyTorch re-export: the graphs are the published FP32 graphs with their weights rounded once (provenance/convert_fp16.py).

	FP32 release	This package
Graphs and external data	1,900,356,386 bytes	950,872,204 bytes
Largest single buffer (token embedding)	256 MiB	128 MiB

ONNX Runtime constant-folds the other weight casts when it creates the sessions, so expect GPU memory for those weights to stay close to the FP32 package. The download and the browser cache halve.

Parity with the FP32 release

The same runtime code (WebBrain's src/chrome/src/providers/d1-runtime.js and its preprocessing) ran both packages on the balanced-V3 held-out test (72 screenshots, 144 questions): on the CPU with onnxruntime-node 1.24.1, and inside the packed WebBrain extension on Chrome WebGPU. No completion probability crosses WebBrain's 0.9 acceptance threshold between the two packages.

Comparison	Decision agreement	Max probability difference	Max expected-score difference
CPU (onnxruntime-node), FP32 vs FP16	144/144	0.0102 (median 5.6e-7, 95th percentile 0.0020)	0.0178
Chrome WebGPU through the WebBrain extension, RTX 5090, FP32 vs FP16	144/144	0.0101	0.0176

Quality on that test is the same as the FP32 release:

Package	Categorical	Completion found	False completions	Score MAE
FP32 release, CPU	92/126	22/24	1/48	0.5742
This package, CPU	92/126	22/24	1/48	0.5728
FP32 release, Chrome WebGPU (RTX 5090)	92/126	22/24	1/48	0.5742
This package, Chrome WebGPU (RTX 5090)	92/126	22/24	1/48	0.5728

These are synthetic, AI-labelled screens, not an independent real-site benchmark. See the FP32 model card for the training history, data limitations and the old test suite on which every stage missed all seven positive completion examples.

On the RTX 5090 the extension downloaded and loaded this package in 85 s and answered a request (one screenshot, two questions) in 0.75 s at the median and 0.88 s at the 95th percentile.

Runtime notes

• The encoder's full-attention layers are dense: each one materializes a 16 × length × length FP32 score tensor per row. WebBrain bounds batches by that tensor (256 MiB, and never more than the adapter's storage-binding limit) and caps text-only rows at 2,048 tokens on WebGPU, truncating the state as the source does at its context limit.
• Vision and projector outputs can stay on the GPU between graphs (preferredOutputLocation: 'gpu-buffer'), see runtime/d1-runtime.js.
• CPU/WASM control and shape operators run as in the FP32 release; this is not a pure-GPU claim.

License and notices

The model is governed by the exact upstream LFM Open License v1.0, not Apache/MIT. Its commercial-use provisions include annual-revenue threshold terms; determine eligibility and obtain any required separate license before commercial deployment. This card is not legal approval or an endorsement by Liquid AI. NOTICE and the per-file sidecars describe the modifications. The copied runtime source carries the WebBrain project license in licenses/webbrain/LICENSE.