Merge recipe

• Attention tensors: t = 0.3
• MLP / expert tensors: t = 0.5
• Other tensors: t = 0.4
• Initial merged checkpoint size before pruning: about 555G

Pruning / compaction

The uploaded safetensors are the compacted pruned checkpoint:

• SnapPrune deep / REAP-style expert pruning
• Prune ratio: 0.3
• Routed experts compacted from sparse original expert ids to dense expert ids
• Final routed experts per MoE layer: 268
• Safetensors checkpoint size: about 394G

License

This checkpoint follows the Modified MIT License from Moonshot AI. See LICENSE.

Commercial attribution requirement: if the Software or derivative works are used for commercial products or services with more than 100 million monthly active users, or more than 20 million US dollars (or equivalent in other currencies) in monthly revenue, you must prominently display Kimi K2.7 Code on the user interface of such product or service.

Status

This model is not fully evaluated yet. Treat it as a research artifact.

At upload time:

• Merge completed successfully.
• Pruning completed successfully.
• Safetensors compaction completed successfully.
• GGUF conversion / quantization / smoke testing may still be in progress.

Use at your own risk and validate quality before production use.