Skip to content
GitHub
View on GitHub

Kimi-K2.5

API reference for Kimi_K2_5

from modal_training_gym.common.models.kimi_k2_5 import Kimi_K2_5

Kimi K2.5 from Moonshot AI.

Inherits from: HFModelConfiguration, ModelConfiguration

FieldTypeDefaultDescription
model_namestr"moonshotai/Kimi-K2.5"HuggingFace repo ID or other model identifier. Default "".
model_path`strNone`"/checkpoints/Kimi-K2.5-bf16"
architecture`ModelArchitectureNone`None
training`ModelTrainingConfigNone`ModelTrainingConfig(gpu_type='H100', n_nodes=4, tensor_model_parallel_size=1, pipeline_model_parallel_size=1, context_parallel_size=1, sequence_parallel=False, expert_model_parallel_size=1, moe_permute_fusion=False, moe_grouped_gemm=False, moe_shared_expert_overlap=False, moe_aux_loss_coeff=0.0, lora_rank=8, lora_alpha=16, target_modules='all-linear', merge_lora=False)

Download or materialize weights into the model volume.

Source: modal_training_gym/common/models/kimi_k2_5.py