Kimi-K2.5-mic

Slim model weights, executable usage and framework expectations in one repository.

• Contract manifest
• Transformers usage
• Transformers expectations
• Fixture provenance
• How the fixture was shrunk

Contract

Kimi-K2.5 vision-language conditional generation

Runnable usage and the framework expectations directory are linked above.

Slim model

Slim fixture: Kimi-K2.5-mic

A 21.5M-parameter fixture for the Kimi-K2-5 MIC, covering moonshotai/Kimi-K2.5. It is a test artifact, not a usable model, tagged test-artifact to keep it out of candidate-model search.

• Fixture key 18c52a9f071d. It covers every checkpoint whose processor files are byte-identical and whose config matches after shrinking.
• Same architecture, smaller config: Kimi_K25ForConditionalGeneration, shrunk by the generic rule in sweep.py (text num_hidden_layers 2, vision num_hidden_layers 2).
• Processor files copied byte for byte from moonshotai/Kimi-K2.5: LICENSE, THIRD_PARTY_NOTICES.md, chat_template.jinja, configuration_deepseek.py, configuration_kimi_k25.py, generation_config.json, kimi_k25_processor.py, kimi_k25_vision_processing.py, media_utils.py, modeling_deepseek.py, modeling_kimi_k25.py, preprocessor_config.json, tiktoken.model, tokenization_kimi.py, tokenizer_config.json, tool_declaration_ts.py.
• Random weights (seed 17). They passed the input-sensitivity gate: every perturbation of the contract's inputs moved the logits by at least 0.001 (weakest: pixel_values:halve, 0.302), so no training was needed.
• Same tensor names as the real checkpoint. The real checkpoint lacks model.language_model.layers.1.mlp.experts.down_proj, model.language_model.layers.1.mlp.experts.gate_up_proj, model.language_model.layers.10.mlp.experts.down_proj, model.language_model.layers.10.mlp.experts.gate_up_proj, model.language_model.layers.11.mlp.experts.down_proj, model.language_model.layers.11.mlp.experts.gate_up_proj, model.language_model.layers.12.mlp.experts.down_proj, model.language_model.layers.12.mlp.experts.gate_up_proj, model.language_model.layers.13.mlp.experts.down_proj, model.language_model.layers.13.mlp.experts.gate_up_proj, model.language_model.layers.14.mlp.experts.down_proj, model.language_model.layers.14.mlp.experts.gate_up_proj, model.language_model.layers.15.mlp.experts.down_proj, model.language_model.layers.15.mlp.experts.gate_up_proj, model.language_model.layers.16.mlp.experts.down_proj, model.language_model.layers.16.mlp.experts.gate_up_proj, model.language_model.layers.17.mlp.experts.down_proj, model.language_model.layers.17.mlp.experts.gate_up_proj, model.language_model.layers.18.mlp.experts.down_proj, model.language_model.layers.18.mlp.experts.gate_up_proj, model.language_model.layers.19.mlp.experts.down_proj, model.language_model.layers.19.mlp.experts.gate_up_proj, model.language_model.layers.2.mlp.experts.down_proj, model.language_model.layers.2.mlp.experts.gate_up_proj, model.language_model.layers.20.mlp.experts.down_proj, model.language_model.layers.20.mlp.experts.gate_up_proj, model.language_model.layers.21.mlp.experts.down_proj, model.language_model.layers.21.mlp.experts.gate_up_proj, model.language_model.layers.22.mlp.experts.down_proj, model.language_model.layers.22.mlp.experts.gate_up_proj, model.language_model.layers.23.mlp.experts.down_proj, model.language_model.layers.23.mlp.experts.gate_up_proj, model.language_model.layers.24.mlp.experts.down_proj, model.language_model.layers.24.mlp.experts.gate_up_proj, model.language_model.layers.25.mlp.experts.down_proj, model.language_model.layers.25.mlp.experts.gate_up_proj, model.language_model.layers.26.mlp.experts.down_proj, model.language_model.layers.26.mlp.experts.gate_up_proj, model.language_model.layers.27.mlp.experts.down_proj, model.language_model.layers.27.mlp.experts.gate_up_proj, model.language_model.layers.28.mlp.experts.down_proj, model.language_model.layers.28.mlp.experts.gate_up_proj, model.language_model.layers.29.mlp.experts.down_proj, model.language_model.layers.29.mlp.experts.gate_up_proj, model.language_model.layers.3.mlp.experts.down_proj, model.language_model.layers.3.mlp.experts.gate_up_proj, model.language_model.layers.30.mlp.experts.down_proj, model.language_model.layers.30.mlp.experts.gate_up_proj, model.language_model.layers.31.mlp.experts.down_proj, model.language_model.layers.31.mlp.experts.gate_up_proj, model.language_model.layers.32.mlp.experts.down_proj, model.language_model.layers.32.mlp.experts.gate_up_proj, model.language_model.layers.33.mlp.experts.down_proj, model.language_model.layers.33.mlp.experts.gate_up_proj, model.language_model.layers.34.mlp.experts.down_proj, model.language_model.layers.34.mlp.experts.gate_up_proj, model.language_model.layers.35.mlp.experts.down_proj, model.language_model.layers.35.mlp.experts.gate_up_proj, model.language_model.layers.36.mlp.experts.down_proj, model.language_model.layers.36.mlp.experts.gate_up_proj, model.language_model.layers.37.mlp.experts.down_proj, model.language_model.layers.37.mlp.experts.gate_up_proj, model.language_model.layers.38.mlp.experts.down_proj, model.language_model.layers.38.mlp.experts.gate_up_proj, model.language_model.layers.39.mlp.experts.down_proj, model.language_model.layers.39.mlp.experts.gate_up_proj, model.language_model.layers.4.mlp.experts.down_proj, model.language_model.layers.4.mlp.experts.gate_up_proj, model.language_model.layers.40.mlp.experts.down_proj, model.language_model.layers.40.mlp.experts.gate_up_proj, model.language_model.layers.41.mlp.experts.down_proj, model.language_model.layers.41.mlp.experts.gate_up_proj, model.language_model.layers.42.mlp.experts.down_proj, model.language_model.layers.42.mlp.experts.gate_up_proj, model.language_model.layers.43.mlp.experts.down_proj, model.language_model.layers.43.mlp.experts.gate_up_proj, model.language_model.layers.44.mlp.experts.down_proj, model.language_model.layers.44.mlp.experts.gate_up_proj, model.language_model.layers.45.mlp.experts.down_proj, model.language_model.layers.45.mlp.experts.gate_up_proj, model.language_model.layers.46.mlp.experts.down_proj, model.language_model.layers.46.mlp.experts.gate_up_proj, model.language_model.layers.47.mlp.experts.down_proj, model.language_model.layers.47.mlp.experts.gate_up_proj, model.language_model.layers.48.mlp.experts.down_proj, model.language_model.layers.48.mlp.experts.gate_up_proj, model.language_model.layers.49.mlp.experts.down_proj, model.language_model.layers.49.mlp.experts.gate_up_proj, model.language_model.layers.5.mlp.experts.down_proj, model.language_model.layers.5.mlp.experts.gate_up_proj, model.language_model.layers.50.mlp.experts.down_proj, model.language_model.layers.50.mlp.experts.gate_up_proj, model.language_model.layers.51.mlp.experts.down_proj, model.language_model.layers.51.mlp.experts.gate_up_proj, model.language_model.layers.52.mlp.experts.down_proj, model.language_model.layers.52.mlp.experts.gate_up_proj, model.language_model.layers.53.mlp.experts.down_proj, model.language_model.layers.53.mlp.experts.gate_up_proj, model.language_model.layers.54.mlp.experts.down_proj, model.language_model.layers.54.mlp.experts.gate_up_proj, model.language_model.layers.55.mlp.experts.down_proj, model.language_model.layers.55.mlp.experts.gate_up_proj, model.language_model.layers.56.mlp.experts.down_proj, model.language_model.layers.56.mlp.experts.gate_up_proj, model.language_model.layers.57.mlp.experts.down_proj, model.language_model.layers.57.mlp.experts.gate_up_proj, model.language_model.layers.58.mlp.experts.down_proj, model.language_model.layers.58.mlp.experts.gate_up_proj, model.language_model.layers.59.mlp.experts.down_proj, model.language_model.layers.59.mlp.experts.gate_up_proj, model.language_model.layers.6.mlp.experts.down_proj, model.language_model.layers.6.mlp.experts.gate_up_proj, model.language_model.layers.60.mlp.experts.down_proj, model.language_model.layers.60.mlp.experts.gate_up_proj, model.language_model.layers.7.mlp.experts.down_proj, model.language_model.layers.7.mlp.experts.gate_up_proj, model.language_model.layers.8.mlp.experts.down_proj, model.language_model.layers.8.mlp.experts.gate_up_proj, model.language_model.layers.9.mlp.experts.down_proj, model.language_model.layers.9.mlp.experts.gate_up_proj, model.vision_tower.layers.0.mlp.fc2.weight, model.vision_tower.layers.1.norm2.weight, model.vision_tower.layers.16.mlp.fc2.bias, model.vision_tower.layers.16.norm2.weight, model.vision_tower.layers.21.norm2.weight, so the fixture omits it too, and loading initializes it as it does for users.
• Saved in bfloat16, the dtype moonshotai/Kimi-K2.5 loads in by default.
• Finite in float32, bfloat16, and float16 (forward, loss, gradients).

Built by build_fixtures.py; build.json records the pinned sources, gate results, and file hashes. This repository includes the slim weights, usage and framework expectations; rebuild the other files with python example/build_fixtures.py --fixture Kimi-K2.5-mic.