Downloads · 30 days
25
31% of all-time downloads
Kimi-K2.5-BF16-Pruned-20B-FP8
ubicloud/Kimi-K2.5-BF16-Pruned-20B-FP8
Kimi-K2.5-BF16-Pruned-20B-FP8 is a machine learning model from ubicloud. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
This model is a pruned variant of ananayarora/Kimi-K2.5-BF16 that retains the first 2 layer(s) of the original 61 layer(s) architecture. It is intended for pipeline testing and performance research rather than product…
All-time downloads
80
Public
Parameters
20.4B
20.8 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors20.8 GB · 100%
Parameter types
How the weights are stored.
F8_E4M320B · 98%
Base models
Model card
From the Hugging Face model README
This model is a pruned variant of ananayarora/Kimi-K2.5-BF16 that retains the first 2 layer(s) of the original 61 layer(s) architecture. It is intended for pipeline testing and performance research rather than production use.
Made with ❤️ by Model Pruner
Kimi K2.5 — BF16 Safetensors
BF16 (bfloat16) conversion of Moonshot AI's Kimi K2.5.
Converted from the official native INT4 weights so that LlamaFactory + KTransformers can run LoRA SFT directly without a per-run conversion step.
| Spec | Value |
|---|---|
| Source | moonshotai/Kimi-K2.5 |
| Format | safetensors (BF16) |
| Shards | 64 |
| Total size | ~1913 GB |
| Architecture | MoE — 1T total, 32B active |
| Context | 256K tokens |
Usage with LlamaFactory + KTransformers
model_name_or_path: ananayarora/Kimi-K2.5-BF16
stage: sft
finetuning_type: lora
bf16: true
use_kt: true
License
Same as the base model: Modified MIT.
- Type
- kimi_k25
- Created
- Mar 30, 2026
- Updated
- Mar 30, 2026