Kairos-Initialized
⚠️ Experimental research artifact — NOT competent. This checkpoint comes from a series of experiments on how to build a small multimodal model from scratch. It is not a usable vision-language model. Do not use it in production, for automated decisions, or in any safety-critical context.
Overview
Kairos-Initialized is the assembled-but-untrained scaffold of Kairos, a research project exploring how to build a small multimodal model by combining an existing vision encoder, a lightweight projector, and a small LLM.
- Vision tower: Aquiles-ai/MoonViT-3D, extracted from moonshotai/Kimi-K2.6. Frozen in every training stage.
- Projector: Kimi-style 2-layer MLP (
LayerNorm → Linear(4608→4608) → GELU → Linear(4608→2048)), randomly initialized, no alignment training applied. - LLM: LiquidAI/LFM2.5-2.6B.
- Modality: image ↔ text only (no video).
The architecture, projector and processor follow the Kimi-K2.5/Kimi-K2.6 implementation on transformers 5.x.
Behavior
- Text-only input: behaves like the base LFM2.5-2.6B. The projector is never touched, so the text-only prior is intact.
- Image input: the random projector emits embeddings roughly 65× the norm of real text embeddings (no output cap is configured in this checkpoint). That injects disproportionately large vectors into a frozen LLM, so image-grounded output is garbage and unrelated to the actual image content.
Purpose
This checkpoint is the starting point of the training pipeline. It is used to:
- Validate the plumbing (special tokens, embedding dimensions, forward/backward flow, frozen backbone).
- Serve as the seed for the next experiment stages.
It exists because it was produced by the experiments, not because it is a usable model. See the progression:
| Checkpoint | What changed | Status |
|---|---|---|
| Kairos-Initialized | Assembled architecture, random projector | Scaffold only |
| Kairos-Proj-80k | Projector aligned on 80k image-caption pairs (LLM frozen) | Not a usable VLM |
| Kairos-Alig-30k | Projector + LoRA(LLM) trained together on 30k reasoning samples | Early experiments, still not competent |
Requirements
transformers>= 5.x (developed with 5.14.1)trust_remote_code=True(the modeling code ships with the checkpoint)
Usage
import torch
from transformers import AutoModelForCausalLM, AutoProcessor
model_id = "Aquiles-ai/Kairos-Initialized"
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True, dtype=torch.bfloat16)
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
messages = [{"role": "user", "content": "Hello!"}]
enc = processor.apply_chat_template(messages, tokenize=True, return_dict=True, return_tensors="pt", add_generation_prompt=True)
out = model.generate(**enc, max_new_tokens=64)
Limitations
- Not competent for image understanding. Random projector, zero alignment.
- Image inputs produce out-of-distribution embeddings (explosive norms) and incoherent output.
- Only meaningful as the base scaffold of an experiment pipeline.
References
- Blog post: Kairos: Building a Multimodal Model with LFM2.5 and Kimi-K2.6
- Public repo: Aquiles-ai/Kairos
- Training dataset: Aquiles-ai/LLaVA-CC3M-Pretrain-595K-Embedded
- Reasoning dataset: Aquiles-ai/Kairos-Multimodal-Reasoning
License
This model card describes a research artifact assembled from third-party components (MoonViT-3D from Kimi-K2.6, LFM2.5-2.6B). Check each component's license before any use.