Course 0 — Start Here: The K3 Thesis

Kimi K3 is easiest to understand as a claim about two kinds of scale. Pretraining scale changes the policy and representations available before a task begins. Test-time scale changes how much reasoning, tool use, environment interaction, and verification that policy can spend on a particular task. They are complementary; neither parameter count nor a long answer alone establishes intelligence.

Two axes, not one

Imagine a coordinate system:

• the horizontal axis is learned capacity produced by data, model architecture, and pretraining compute;
• the vertical axis is useful inference-time effort produced by reasoning budgets, tools, environments, and verification.

K3's paper argues for movement on both axes. Sections 2–3 build the foundation. Sections 4–5 make extended agentic work trainable and operable. Sections 6–7 measure the resulting system and show selected trajectories. This diagram is a learning model, not a paper-reported ablation proving the contribution of each axis.

Four numbers that mean different things

The paper reports 2.8 trillion total parameters, 104 billion active parameters, and a 1 million-token context window, alongside native vision. These quantities must not be collapsed:

• total parameters describe available sparse capacity;
• active parameters describe the subset used for a token and are not identical to FLOPs;
• context length is an addressable capacity, not proof of reliable use at every distance;
• native vision means visual features are trained into the shared causal system, not that raw pixels are text tokens.

All four are paper-reported properties. This repository does not reproduce a 2.8T model or full-scale training run.

The paper as one dependency map

1. Architecture (§2): route information across sequence, depth, width, and modality.
2. Pretraining (§3): construct data, allocate compute, stabilize training, and extend context.
3. Post-training (§4): train effort levels, agentic environments, distillation, and deployment-aware behavior.
4. Infrastructure (§5): make sparse training, long-context RL, and hybrid serving feasible.
5. Evaluation (§6): measure a capability profile under explicit protocols.
6. Case studies (§7): inspect selected end-to-end trajectories and artifacts.
7. Conclusion (§8): state what the combined evidence supports and where gaps remain.

The evidence contract

Every course surface uses one of three labels:

Label	What it can establish
Executed here	A deterministic miniature, assertion, or retained output ran in this repository.
Animated interpretation	Motion explains a mechanism or relationship; it is not a measurement.
Reported by paper	The full-scale claim comes from Kimi K3 and has not been independently reproduced here.

Source: Author-created evidence taxonomy for reading Kimi K3; not a paper table.

Use the course in this rhythm: watch the mechanism → read the derivation → inspect the notebook → check the evidence boundary. Start with the architecture course, then follow the numbered documentation chapters.

Before moving on

Explain, in one sentence each, why total parameters are not active parameters, why context capacity is not demonstrated retention, and why an animated mechanism is not experimental evidence. If any distinction is unclear, revisit the table above before beginning Course 1.