Kimi K3 Q2_K-to-t158 MTPLX

This is the ordinary, fully resident Safetensors serialization of the experimental Kimi K3 MTPLX t158 checkpoint. The SSD-streaming serialization is published separately at OpensourceWTF/Kimi-K3-Q2_K-t158-MTPLX-streaming.

[!IMPORTANT] This is not Moonshot AI's original Kimi K3 checkpoint, and it is not a llama.cpp/GGUF model. The 92 routed MoE layers were requantized from GrEarl/Kimi-K3-GGUF Q2_K to the MTPLX t158 codec. No downstream quality evaluation has been completed.

File layout

This repository is one indexed Safetensors checkpoint:

• model-00001-of-00188.safetensors through model-00096-of-00188.safetensors: resident model tensors
• model-00097-of-00188.safetensors through model-00188-of-00188.safetensors: one routed expert layer per shard
• model.safetensors.index.json: all 3,180 tensor-to-shard mappings
• mtplx_t158.json: the packed-expert serialization contract

Each routed layer is represented by six valid Safetensors tensors: gate_proj.packed, gate_proj.scales, up_proj.packed, up_proj.scales, down_proj.packed, and down_proj.scales. Expert ID is the leading dimension. The packed weights use Safetensors U8, and the BF16 scale bit patterns use U16.

There are no streaming .bin banks, streaming manifests, conversion receipts, or build journals in this repository.

Artifact contract

• Source GGUF: GrEarl/Kimi-K3-GGUF@0169245d3ea1473a3f9f03bca821d855df5fb2a3
• Official metadata: moonshotai/Kimi-K3@9f62e4e9fffbd0a83ddd60e1c209d828994b3569
• Routed topology: 92 MoE layers, 896 experts per layer, top 16
• Expert codec: MTPLX t158, group size 64, 1.875 physical bits per weight
• Expert tensor payload: 638,142,382,080 bytes
• Resident tensor payload: 114,404,258,816 bytes
• Indexed tensor payload: 752,546,640,896 bytes
• Source manifest SHA-256: 5dc510958199ad6b9d5701e47dd96bff14c41fbec1531a7a93ed16dc0176bcc1

The normal shards are a deterministic byte reordering of the verified streaming expert banks. Repacking does not decode or requantize values. Every one of the 82,432 source expert-record SHA-256 digests is checked during the conversion, and every completed shard receives a full readback hash before it is installed.

Runtime compatibility

Safetensors has no native t158 dtype. These are standards-compliant Safetensors containers using packed U8 and scale-bit U16 tensors, but the expert tensors require an MTPLX t158-aware loader. A generic Transformers from_pretrained call does not know how to execute this packed expert layout.

The normal checkpoint's indexed tensor payload is about 752.5 GB, so it was structurally and cryptographically verified rather than loaded fully on the 512 GB conversion host. For the tested bounded-memory runtime, use the separate SSD-streaming repository.

Scope and quality

The current MTPLX K3 runtime is autoregressive, text-only, and does not include MTP. The retained serialization pilot measured cosine 0.903718 against the decoded Q2_K values for one selected expert. That is a narrow conversion diagnostic, not a model-quality score or benchmark.

Kimi K3 is released under the Kimi K3 License. See the moonshotai/Kimi-K3 model card for the original architecture, evaluation, and usage documentation.