Ling
1 MLX alias · inclusionAI's Ling-3.0-tiny · 7.9B MoE with 1.3B active, KDA+MLA hybrid, MIT.
inclusionAI's Ling 3.0 line — sparse-MoE reasoners on a KDA + MLA hybrid attention stack (three linear-attention layers per full-attention layer), MIT licensed. Ling-3.0-tiny is 7.9B parameters total with only 1.3B active per token (128 experts, top-8 plus one shared), which is how a 131K-context reasoner fits and runs on an 8 GB Mac. Since 0.12.10, rapid-mlx serves it through its own vendored bailing_hybrid implementation — verified against the official modeling code — with thinking streamed as reasoning_content and native tool calling.
- family
- Ling
- aliases
- 1
- lines
- 1
- install
- rapid-mlx serve <alias>
- OpenAI base URL
- http://localhost:8000/v1
Ling 3.0 tiny · 1 alias
7.9B-total / 1.3B-active MoE with the KDA + MLA hybrid stack and 131K context. Uses the glm47 tool-call parser and the qwen3 reasoning parser; thinking toggles via chat_template_kwargs: {"enable_thinking": true} or the model's detailed thinking on/off system-prompt switch.
parser: glm47
| alias | hf repo | tool parser | reasoning | flags |
|---|---|---|---|---|
| ling-3.0-tiny-4bit | rapid-mlx/Ling-3.0-tiny-MLX-4bit | glm47 | qwen3 | hybrid · moe |
Notes & caveats
- ling-3.0-tiny-4bit (4.2 GB) is the pick — the strongest reasoner that fits an 8 GB Mac, mirrored on our edge CDN.
- First MLX conversion of this family — published under the rapid-mlx org on Hugging Face.