status: not 1:1 speed yet. currently served via fallback engine; expect far below dense-model tok/s. fast-path fix in progress.

rhm-qwythos-9b-caix
A Core AI (.aimodel) bundle of empero-ai/Qwythos-9B-Claude-Mythos-5-1M, converted for
native on-device inference on Apple silicon with caix.
Verified: loads through caix and returns visible OpenAI message.content on Apple silicon.
| Base model | empero-ai/Qwythos-9B-Claude-Mythos-5-1M |
| Format | Apple Core AI .aimodel |
| Architecture | qwen3_5 hybrid linear/full attention |
| Compression | 4bit |
| Max context | 4096 |
| Size | ~4.7 GB |
| License | Apache-2.0 |
Use with caix
hf download redhillsmediafl/rhm-qwythos-9b-caix \
--local-dir models/exports/qwythos-9b-coreai
./caix serve
Runs 100% on-device. Requires a Core AI-capable macOS beta and caix.
Verification
Converted from the official Empero AI release, then checked with caix's OpenAI-compatible API:
curl http://127.0.0.1:1240/v1/chat/completions \
-H 'content-type: application/json' \
-d '{"model":"qwythos-9b-coreai","messages":[{"role":"user","content":"Answer with only the city name: What is the capital of France?"}],"max_tokens":16,"temperature":0}'
Expected visible output: Paris.
Conversion Note
Qwythos's published chat template exposes a no-thinking generation branch behind enable_thinking=false.
Apple's CoreAILanguageModels tokenizer API currently accepts messages and tools, but not arbitrary template
kwargs, so caix postprocesses this exported tokenizer template to bake in that no-thinking branch. Without
that patch, generic OpenAI clients receive hidden reasoning_content with empty visible content.
License
The base/source model is licensed Apache-2.0; this Core AI re-packaging inherits that license. See the base model card for the full notice.
Beta. caix and Apple Core AI are beta software. Converted from the official Qwythos release. Unaffiliated with Apple.
Support caix
caix is free and open. No paywall.
- Test requests: https://github.com/RedHillsMediaFL/caix/blob/main/docs/TESTER_REQUESTS.md
- Support packs: https://redhillsmediafl.com/open-source
- Shop: https://redhillsmediafl.com/shop
Support is optional and never gates features.
More open-source work: redhillsmediafl.com/open-source.
Model tree for redhillsmediafl/rhm-qwythos-9b-caix
Base model
Qwen/Qwen3.5-9B-Base