H2O-Lightning-4B ONNX (q4 / q4f16)
ONNX weights of h2oai/h2o-lightning-4b (revision v1.2.3) for transformers.js and onnxruntime-web. The graphs are those of onnx-community/Qwen3.5-4B-ONNX, with the fine-tuned weights swapped in and re-quantized (MatMulNBits, 4 bit, block 32). The vision encoder is unchanged from the base model, because the fine-tune did not touch it. The decoder computes logits for the last position only.
Live demo: mrfakename/h2o-lightning-4b-webgpu
Parity with the bf16/fp32 PyTorch model
On 6 decision prompts (label probabilities read after the Answer: prefill, as in the model's serving shim):
- Top label: same as PyTorch on 5 of 6, for both q4 and q4f16. The miss is a borderline yes/no, where PyTorch gives P(yes) 0.60 and ONNX gives 0.44.
- Probability difference: median 0.02, max 0.16.
- Greedy free text: the first 4 to 5 tokens match.
See parity_report.json.
This is a decision model. Use the prompt format and label readout described on the original model card.