kev-4b-ONNX

kev-4b (revision 139fdd94f1b6a6ad80cc15e08fcb99cac885a101) by Jared Palmer, packaged for the browser by runonweb (runonweb/classify).

Kev is a Jev-style decision model: typed questions about one text (noul yes/no, choice, score) answered with calibrated probabilities, no text generation. Requests follow TypeSafe's System One API.

What changed from the original

• The rank-16 LoRA is merged into Qwen/Qwen3.5-4B-Base (revision 1001bb4d826a52d1f399e183466143f4da7b741b) in fp32.
• The backbone is exported with the onnxruntime-genai builder without the LM head and MTP layer: the graph returns hidden_states plus the recurrent/conv/KV cache, so the state is encoded once and each question runs as its own row on that cache (Kev's row form).
• Weights are int4 (RTN, block 32), embeddings included; activations fp32. ~2.7 GB, split in two files under 2 GB. Needs the LinearAttention / CausalConvWithState contrib ops: ONNX Runtime Web's native WebGPU build (onnxruntime-web/webgpu) on the GPU, onnxruntime-web/wasm on the CPU.
• The pointer head is head.bin (fp32: q.weight, q.bias, k.weight, k.bias); kev.json holds the calibration temperature, delimiter token ids and cache layout.

On 192 questions from Kev's development suites, probabilities differ from the fp32 export by 0.040 on average; 10 answers change, 3 with a margin above 0.2 (accuracy 0.776 vs 0.766). The fp32 export of the same graph matches Kev's PyTorch fp32 path within 4e-5. Recipe: training/kev-onnx in the runonweb repo.

Use

import { Classifier } from 'runonweb/classify'

const classifier = new Classifier({ model: 'midudev/kev-4b-ONNX' })
const { answers } = await classifier.classify({
  state: 'I was charged twice. Please fix this ASAP.',
  questions: { billing: { type: 'noul', instructions: 'Is this ticket about billing?' } },
})

License

Apache-2.0, like Kev and the Qwen3.5 base. Kev's training datasets have their own licenses; see the original model card.