iKev — Kev-0.8B for MLX (8-bit)

Kev decision model (Qwen3.5-0.8B-Base + Kev LoRA, merged) with Kev's pointer head, converted for mlx-swift-lm. Loaded by the iKev Swift package.

Kev answers fixed-choice questions about text in one prefill pass — no generation, temperature-calibrated probabilities, a confidence you can threshold on.

Files

• model.safetensors, config.json — merged backbone, MLX 8-bit (group size 64)
• head.safetensors — pointer head (q/k Linear 1024→256)
• kev.json — format metadata and fitted temperature
• tokenizer files

Usage

import Kev
let model = try await KevModel.load(hubID: "gtokman/iKev")
let answers = try await model.decide(state: "set a timer for 10 minutes",
                                     questions: [.choice("What does the user want?", options: ["casual", "deviceAction", "assistant"])])

Parity vs Kev's Python MLX backend: max |Δp| ≈ 0.02 on the reference set. Produced with scripts/convert.py --bits 8 from the iKev repo.