---
license: mit
library_name: llama.cpp
pipeline_tag: text-generation
base_model:
  - inference-optimization/Kimi-K3-0.40B
  - inference-optimization/Kimi-K3-0.40B-MXFP4
tags:
  - gguf
  - llama.cpp
  - kimi-k3
  - mxfp4
  - moe
  - test-model
---

# Kimi-K3 0.40B GGUF validation fixtures

GGUF conversions of the tiny Kimi-K3 0.40B development checkpoints.

These files are intended for:

- Kimi-K3 architecture validation;
- `llama.cpp` conversion and inference testing;
- MXFP4 MoE layout testing;
- out-of-core expert-loading research.

They are **not intended as useful language models** and should not be treated as
quality proxies for the full Kimi-K3 model.

## Files

### `Kimi-K3-0.40B-F16.gguf`

Reference F16 conversion of:

`inference-optimization/Kimi-K3-0.40B`

### `Kimi-K3-0.40B-MXFP4.gguf`

Hybrid conversion of:

`inference-optimization/Kimi-K3-0.40B-MXFP4`

Representation:

- 168 routed expert projection tensors remain in MXFP4;
- they are repacked into 21 GGUF expert groups;
- 35 resident MoE tensors are dequantized to F16;
- unknown packed tensors are rejected by the converter.

## Converter

Converted with:

- repository: `murillo128/llama.cpp`
- branch: `k3/out-of-core`
- converter commit: `84245db4c790af22135f34992689edcc11877003`

See `conversion-manifest.json` for:

- exact source revisions;
- generated file sizes;
- SHA-256 checksums;
- conversion counts;
- Python package versions.

## Smoke-test prompt

```text
According to all known laws
```

Both F16 and hybrid MXFP4 fixtures were loaded and executed successfully using
the CPU backend. The observed fixture continuation was:

the start.

This is only a loading and execution smoke test, not a correctness or quality
benchmark.

Project

These artifacts support the k3-out-of-core project, which explores tiered
resident, RAM, and NVMe storage for routed MoE experts in llama.cpp.