kev.js weights

Browser-ready exports of Kev, Jared Palmer's family of small decision models. They answer yes/no, multiple-choice and rating questions with calibrated probabilities, and run in a browser on WebGPU through @ai-ecoverse/kev.js. This repo holds only converted weights: no training, evaluation or model design here is ours.

import * as ort from "onnxruntime-web/webgpu";
import { loadKev } from "@ai-ecoverse/kev.js";

const kev = await loadKev("https://huggingface.co/ai-ecoverse/kev.js/resolve/main/kev-0.8b", { ort, variant: "q8f32" });
const res = await kev.systemOne({
  state: "I was charged twice. Please fix this ASAP.",
  questions: { billing: { type: "noul", instructions: "Is this ticket about billing?" } },
});

Contents

Folder	Variant	Download	Base	Source checkpoint
kev-0.8b	q8f32	0.84 GB	Qwen/Qwen3.5-0.8B-Base	jaredpalmer/kev-0.8b@9a45d25eb2ab761841196625383fa1dff0e56c1e
kev-0.8b	q8	0.80 GB	Qwen/Qwen3.5-0.8B-Base	jaredpalmer/kev-0.8b@9a45d25eb2ab761841196625383fa1dff0e56c1e
kev-4b	q8f32	4.69 GB	Qwen/Qwen3.5-4B-Base	jaredpalmer/kev-4b@139fdd94f1b6a6ad80cc15e08fcb99cac885a101
kev-9b	q8f32	8.84 GB	Qwen/Qwen3.5-9B-Base	jaredpalmer/kev-9b@2629c06a5aeb0feb3b9783bafed17ed8f39ecf5c

Every bundle is int8 weights with fp32 activations, split into 32 MB files so any CDN or proxy can serve them. manifest.json lists the files, their sizes, the tokenizer and the pointer head, plus the measured deviation from the original fp32 PyTorch model on a fixture set.

Provenance and licenses

• Models and training: jaredpalmer/kev (Apache-2.0). Source checkpoints: jaredpalmer/kev-0.8b@9a45d25eb2ab761841196625383fa1dff0e56c1e, jaredpalmer/kev-4b@139fdd94f1b6a6ad80cc15e08fcb99cac885a101, jaredpalmer/kev-9b@2629c06a5aeb0feb3b9783bafed17ed8f39ecf5c.
• Base models: Qwen3.5 (Apache-2.0) — via the Kev checkpoints above (Qwen/Qwen3.5-0.8B-Base, Qwen/Qwen3.5-4B-Base, Qwen/Qwen3.5-9B-Base).
• Architecture described in Jev's Architecture Unmasked. The API shapes follow TypeSafe's System One; Jev is TypeSafe's hosted model and is not affiliated with this repo.
• Conversion: LoRA merged in fp32, exported with the onnxruntime-genai model builder without the LM head, embeddings quantized to int8 per row. Details in the kev.js README.
• -vision folders: the same decoder with an image_embeds input, plus the base model's own Qwen3.5 vision tower and patch merger (fp16 weights), unmodified and not trained with Kev. Accuracy on images is in the kev.js README.