KevANE-0.6B

KevANE-0.6B

A Core ML release of Kev 0.6B for Apple Silicon. It answers TypeSafe System One choice, score, and yes/no questions through the KevANE server. This is a decision model, not a chat or text-generation model. It is not directly usable with transformers.pipeline or AutoModelForCausalLM.

What is in this repository?

The merged PyTorch backbone is a build intermediate and is not required for inference. The server and pinned Kev compatibility code are in the GitHub repository.

Run locally

Requires an Apple Silicon Mac. Follow the KevANE installation guide:

git clone https://github.com/LCW0NJUPT/KevANE.git
cd KevANE
bash scripts/install.sh
~/.local/bin/kev-ane start --wait

Example request:

curl --noproxy '*' http://127.0.0.1:8008/v1/systemone \
  -H 'Content-Type: application/json' \
  -d '{"model":"kevane-0.6b","state":"The user says hello.","questions":{"intent":{"type":"choice","instructions":"What is the intent?","criteria":{"greeting":null,"request":null}}}}'

Use ~/.local/bin/kev-ane stop when finished. No manual conversion or compilation step is needed. Core ML prepares the downloaded model for the local Mac on first load and may briefly use a full CPU core. KevANE keeps the result in the ignored source checkout's build/compiled/ directory (roughly another 1.1 GB); subsequent starts reuse it. Run kev-ane cache status to inspect it or kev-ane cache clear while the service is stopped to remove it. The service binds to localhost by default. Jarvis users can set its TypeSafe/System One URL to http://127.0.0.1:8008 and model name to kevane-0.6b, with a non-empty TypeSafe key such as local.

Architecture and provenance

The Kev 0.6B checkpoint is a LoRA adaptation of Qwen3-0.6B-Base. KevANE merges that adaptation, converts the 28-layer backbone to a Core ML ML Program using ANEMLL, and retains Kev's pointer head on CPU. The Core ML graph is FP16, with no LUT quantization and a fixed 512-token packed context. Conversion inputs are token IDs, position IDs, an additive branch mask, and Core ML state; the output is final hidden states.

Pinned revisions and exact input shapes are in kev-qwen3-0.6b-hidden-meta.json. The KevANE runtime includes selected Kev source modules at the pinned revision and does not require a separate Kev installation. The server defaults to CPU_AND_NE, which excludes GPU for the Core ML backbone but does not guarantee ANE placement for every operation. Tokenization and the pointer head use CPU.

Verification and limitations

The recorded FP16/FP32 comparison reports minimum hidden-state cosine similarity 0.999058 across seven samples and matching top answers for 34/35 decisions. The changed answer was a near tie. These are conversion checks on a small fixture set, not a general accuracy benchmark.

The Core ML graph accepts 512 tokens per pass. KevANE can evaluate up to 16 independent questions in separate passes when a packed request exceeds 512 tokens, but shared state plus any single question must still fit or the server returns HTTP 422. No input is silently truncated. This short context can limit decisions based on long chat histories, and multiple passes add latency. Close decisions may change under FP16. The model only scores options supplied in a System One request; it does not compose answers or verify the truth of its input. For device-level ANE attribution, profile the running app with Instruments rather than relying only on the CPU_AND_NE setting.

License

Apache-2.0. The converted weights derive from Kev and Qwen3-0.6B-Base, both released under Apache-2.0. See NOTICE for attribution. This repository is an independent conversion and is not an official release of Kev, Qwen, ANEMLL, or Jarvis.