Inkling 2 Mini: how a looped MoE model runs on a laptop

12 steps · about 6 h 3 min · Intermediate

  1. But what is a GPT? Visual intro to transformers | Chapter 5, Deep Learning (video, 3Blue1Brown (YouTube), 27 min)

    Everything about Inkling 2 Mini builds on the basic transformer, so start with tokens, embeddings and the predict-sample-repeat loop.

  2. Attention in transformers, step-by-step | Deep Learning Chapter 6 (video, 3Blue1Brown (YouTube), 26 min)

    Attention is the core operation inside each layer that the model repeats, so you need it before talking about looping layers.

  3. Neural Networks: Zero To Hero (tutorial, Andrej Karpathy, 2 h)

    Building a small GPT in code (the 'Let's build GPT' lecture) turns the transformer layer stack from an abstract diagram into something you have actually written.

  4. Mixture of Experts Explained (article, Hugging Face Blog, 30 min)

    Inkling 2 Mini is 26B total but only a few billion active parameters per token, and this is the clearest explanation of why.

  5. Mixture of Experts (MoEs) in Transformers (article, Hugging Face Blog, 20 min)

    A newer follow-up covering routers and how experts replace feed-forward layers, which makes the 128-experts, 8-per-token design easier to read.

  6. Looped Transformer (reference, Sebastian Raschka, 10 min)

    A short, clear entry on the key idea behind Mini: reuse the same layers several times to get more computation without more stored weights.

  7. Recursive Transformers (guide, Aman's AI Journal, 35 min)

    Goes deeper on recurrent-depth models, latent reasoning and the chain-of-thought monitorability question that the Mini announcement raises.

  8. Quantization concepts (doc, Hugging Face docs, 20 min)

    Mini was trained at 4-bit precision and ships as small GGUF files, so you need to know how fewer bits map to rounding error.

  9. LLM Quantization Explained: Run Bigger Models on Less RAM (guide, Enclave AI, 15 min)

    Decodes file names like Q4_K_M so you can pick the right download for your RAM.

  10. Introducing Inkling 2 Mini (article, Thinking Machines Lab, 15 min)

    With the concepts in place, the primary source on Mini's layers, training and benchmarks becomes readable.

  11. Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup (blog, Hugging Face Blog, 15 min)

    An independent technical walkthrough of the architecture, with 48 layer visits per token and expert routing, plus how to use the weights.

  12. Inkling 2 Mini - Run Locally Guide (guide, Unsloth docs, 30 min)

    The last step is hands-on: download a quantized build and run it on your own machine with llama.cpp or Unsloth Desktop.