FluidInference/kev-0.6b-coreml

Kev 0.6B Core ML

This is a fixed-shape Core ML conversion of the trained Kev 0.6B LoRA adapter and pointer head merged with Qwen3-0.6B-Base. The source/ files in this repository contain the exact conversion scripts and reproducibility notes.

Artifact

kev_0_6b_fp16_L128_options32.mlpackage/ is the FP16 Core ML package; kev_0_6b_w8_L128_options32.mlpackage/ is the 8-bit weight compressed variant. Both target iOS 17/macOS 14 or later and accept one Kev typed question per call, at most 128 tokens and 32 option slots. Each package contains the merged Qwen3 backbone and trained pointer head. source/runtime.py uses the included tokenizer, pinned serving configuration, and Core ML package without loading the upstream adapter or Qwen weights. The state may be shortened to fit L128, and requests with an overlong question branch fail explicitly. These variants do not implement Kev's native packed multiquestion forward in one call.

Download the repository with snapshot_download("FluidInference/kev-0.6b-coreml"). From its root, install the pinned Python dependencies and run an unlabelled one-question request:

uv sync --project source
uv run --project source python source/runtime.py --model-dir . --request-json request.json

For example, request.json can be {"state":"The customer was charged twice.","questions":{"team":{"type":"choice","instructions":"Which team should handle this?","criteria":{"billing":"Charges and refunds","support":"Technical issues"}}}}. Pass --precision w8 for the compressed package. The runtime returns the upstream typed System One answer. source/verify.py is a separate conversion parity script that intentionally loads the upstream PyTorch model.

The Core ML inputs are input_ids [1,128], attention_mask [1,128], decide_map [1,1,128], and option_map [1,32,128]; outputs are logits and probabilities over 32 slots. Slots absent from the request carry a large negative logit.

Validation and limits

The explicit PyTorch exporter matched the trained Kev model within 9.54e-7 logits on the export fixture. Both Core ML packages selected the same answer on 4/4 local requests covering Choice, Noul, and Score. FP16 had maximum probability difference 0.002031 and size 1,194 MB. W8 had maximum probability difference 0.008554 and size 599 MB. The standalone host passed offline, unlabelled Choice, Noul, and Score requests without native weights; see source/reports/standalone-runtime.json. Short CPU_AND_NE model-call medians on one Apple Silicon Mac ranged 13.10–15.35 ms for FP16 and 12.27–23.15 ms for W8 across two tiny runs; they reversed the speed ordering, so no variant-level speed claim is made. The compute setting permits the Neural Engine but does not establish full ANE residency. These checks do not constitute the Jev Decision Index or a 2048 benchmark, and the tracker’s 31.30 score was measured on a separate serving stack.

The adapter/head and Qwen base are Apache-2.0. The base license text is included as LICENSE. Upstream author: Jared Palmer; Qwen base: Qwen team; Core ML conversion: Fluid Inference. Exact source revisions and trained-weight SHA256 hashes are in source/assets.lock.json.