README.md

# Kimi K3 engineering analysis and Qwen3.8-Max comparison

Human-readable, evidence-backed analysis of Kimi K3's architecture, public
reference code, training system, production optimizations, limitations, and
credible next improvements, plus a release-day comparison with Qwen3.8-Max.

## Start here

Read:

[`analysis/kimi-k3-deep-analysis-en.md`](analysis/kimi-k3-deep-analysis-en.md)

Then read the new comparison:

[`analysis/kimi-k3-vs-qwen3.8-max-en.md`](analysis/kimi-k3-vs-qwen3.8-max-en.md)

The comparison is deliberately provisional. At its 4 August 2026 evidence
snapshot, Qwen3.8-Max was available as a hosted model and its open weights were
promised for the following week, but no official model artifacts were yet found
on the Qwen Hugging Face organization. No paid API testing was performed.

The short French analysis remains available as an earlier summary:
[`analysis/kimi-k3-analysis-fr.md`](analysis/kimi-k3-analysis-fr.md).

## What this bucket contains

- `analysis/kimi-k3-deep-analysis-en.md`: detailed English analysis.
- `analysis/kimi-k3-vs-qwen3.8-max-en.md`: detailed English comparative study.
- `analysis/kimi-k3-analysis-fr.md`: shorter French analysis.
- `data/architecture.json`: structured architecture profile.
- `data/code-audit.json`: structured audit of the public reference code.
- `data/optimization-roadmap.json`: prioritized improvement roadmap.
- `data/kimi-k3-vs-qwen3.8-max.json`: structured comparison snapshot.
- `schemas/*.schema.json`: JSON Schemas for the structured data.
- `schemas/catalog.openapi.yaml`: descriptive read-only OpenAPI contract.
- `diagrams/architecture.mmd`: model architecture.
- `diagrams/training-pipeline.mmd`: agentic training pipeline.
- `diagrams/deployment.mmd`: inference path.
- `diagrams/why-k3-works.mmd`: human mental model of the full system.
- `diagrams/kimi-k3-vs-qwen3.8-max.mmd`: comparison mental model.
- `sources/source-snapshot.json`: analyzed sources and exact revisions.
- `manifest.json`: file sizes and SHA-256 hashes.

## Important scope

This bucket contains no model weights and starts no compute service. It contains
only small text files.

The official weights remain in
[`moonshotai/Kimi-K3`](https://huggingface.co/moonshotai/Kimi-K3). They are
roughly 1.5 TB and require specialized multi-GPU infrastructure for practical
deployment.

The public GitHub repository mainly contains the technical report and
documentation. The Transformers reference code is shipped in the Hugging Face
model repository. That reference code explains the model, but it is not
Moonshot's complete distributed training or production serving stack.

## Bucket access

Public page:

```text
https://huggingface.co/buckets/StephaneGSL/kimi-k3-analysis
```

Bucket URI:

```text
hf://buckets/StephaneGSL/kimi-k3-analysis/
```

Download locally:

```powershell
hf buckets sync hf://buckets/StephaneGSL/kimi-k3-analysis ./kimi-k3-analysis
```

## Main sources

- https://github.com/MoonshotAI/Kimi-K3
- https://huggingface.co/moonshotai/Kimi-K3
- https://github.com/MoonshotAI/MoonEP
- https://github.com/fla-org/flash-linear-attention/pull/691
- https://github.com/kvcache-ai/AgentENV
- https://x.com/Alibaba_Qwen/status/2084100707423289643
- https://github.com/QwenLM/qwen-code
- https://github.com/qwen-code-dev-bot/oh-my-cli