GitHub radar
Ling-3.0-tiny: Fast Reasoning on a MacBook
inclusionAI has announced the release of their new model Ling-3.0-tiny, which is a 7.9B MoE model with a thinking mode running at 86–90 tokens/sec on M4 Pro Macbook and using ~8GB RAM.
Ling-3.0-tiny is a hybrid MoE reasoning model proposed by inclusionAI, containing 7.9B parameters in total but only using 1.3B parameters each token. Using KDA-MLA architecture and 128-expert sparse MoE (8 routed + 1 shared expert active per token), it has high inference speed. Running at 86–90 tokens/sec on MacBook M4 Pro with FP8 mode, consuming about 8.34GB memory within an 8K context window. Using YaRN scaling, we can enlarge the context window to 256K tokens. If we set enable_thinking=true, temperature=1.0, top_p=0.95, it would do some step-by-step reasoning before giving out the final result. You can try it via SGLang(pre-built docker image) or a special vLLM branch or an open pull request of Ollama for apple silicon.
Why a vibe-coder should care
It’s a tiny but fast reasoning model, which runs well on normal macbook. Ideal for scenarios where you need to think step by step but lack fund to pay for cloud computing services. inclusionAI is a Chinese AI group. Note that the pull request for Ollama support hasn’t been merged yet. The most stable way to use this model is by SGLang with docker image.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install Ling-3.0-tiny from inclusionAI for me: https://huggingface.co/inclusionAI/Ling-3.0-tiny — run it with SGLang (Docker) following the model card instructions and show me how to call it with thinking mode enabled (enable_thinking=true).
A regular laptop with 16 GB of RAM or more — in FP8 format the model takes up about 8 GB and delivers 86–90 tokens per second on a MacBook with an M4 Pro chip.
Open on Hugging Face▌ More finds