LESSON 8.7 · Code Study · 180 min
Kimi K3: audit a 3T-class open-weight system
Kimi K3: audit a 3T-class open-weight system. Learn KDA, AttnRes, Stable LatentMoE, MXFP4 QAT, million-token agentic RL, and the Kimi K3 License and complete: Build a five-layer au
Learning objectives
- Draw the sequence-, depth-, and width-wise information paths behind KDA, AttnRes, and Stable LatentMoE.
- Separate 2.8T total parameters from 104B active parameters, then account for storage, communication, KV state, and runtime overhead.
- Audit an evaluation claim by matching reasoning effort, agent harness, tool access, benchmark split, cost, and preserved failures.
- Distinguish open weights from unrestricted open source and identify when the Kimi K3 License requires a separate commercial review.
Core concepts
KDA
Kimi Delta Attention supplies linear recurrent attention for most layers, while periodic Gated MLA layers restore global content-based access. Audit the 3:1 pattern and the final global-attention layer rather than reducing the design to “linear attention”.
AttnRes
Attention Residuals let a block selectively read earlier block outputs and embeddings. Track which depth-wise paths are available and what extra state or communication they require.
Stable LatentMoE
The model routes through 896 experts and selects 16 per token in a compact latent expert space. RMSNorm, SiTU-GLU, and quantile balancing are stabilization mechanisms, not optional footnotes.
MXFP4 QAT
Quantization-aware training begins during post-training and targets MXFP4 expert weights with MXFP8 activations. Do not infer that the entire checkpoint or serving stack occupies four bits per parameter.
Million-token agentic RL
The post-training stack combines SFT, long-horizon agentic RL, preserved thinking history, and multiple effort levels. A one-million-token window is capacity; usable long-horizon behavior still needs trajectory and compaction tests.
Build and verify
Build a five-layer audit matrix spanning architecture, training, inference, evaluation, and licensing
- Predict: write what 2.8T total, 104B active, 896 routed, 16 selected, and a one-million-token context do—and do not—imply.
- Build: give every claim a source, evidence class, measurement boundary, dependency, and falsification test.
- Verify: cross-check the technical report, official repository, and exact license text; label official claims separately from independent evidence.
- Transfer: apply the same matrix to another open-weight model and record which fields cannot be compared directly.