Documentation

Models

Kaiko provides both text models (LLMs) and emotion models. This page explains available models, capabilities, and how to choose the right one.

1. Text Models

Text models generate responses to user messages, similar to other LLM APIs.

Model	Context Window	Typical Latency	Best For
`gpt-4.1-nano`	8k tokens	~200–500ms	Fast, lightweight conversations, real-time agents
`gpt-4.1-small/medium/large`Future	TBD	TBD	Higher accuracy, longer reasoning tasks

Notes

Context window = max combined length of input (prompt_tokens) + output (completion_tokens).
Latency depends on request size and complexity.
Parameters: support standard fields like temperature, max_tokens, top_p, etc.
Tool calling: available in supported models (see Chat API).

2. Emotion Models

Emotion models analyze emotional tone of messages. They return category, text (readable intensity), and raw scores.

Model	Dimensions	Languages	Typical Latency	Best For
`emotion-v1`	joy, sadness, anger, fear, love, surprise	English (multilingual in beta)	~100–200ms	Real-time emotional inference for chat, analytics, coaching bots

Example Response

Emotion Analysis Responsejson

"emotions": {
  "user": {
    "category": "joy",
    "text": "Highly JOY",
    "raw": {
      "joy": 0.99,
      "sadness": 0.002,
      "anger": 0.003
    }
  }
}

3. Model Lifecycle

We regularly update models to improve accuracy and performance.

Stable models (e.g., emotion-v1, gpt-4.1-nano) are production-ready.
Preview models may change without notice (labeled -beta).
Deprecation:
- Announced at least 90 days in advance
- Notice appears in Changelog and API headers
- Sunset dates documented clearly

Example deprecation header:

Deprecation Headersbash

Kaiko-Deprecation: 2025-12-01
Kaiko-Replacement: emotion-v2

4. Choosing a Model

Use text models when you need natural language generation.
Use emotion models when you need affective state detection.
Use both in the Unified Chat API to combine generation + emotion in one call.
For lightweight bots or pipelines, use the Non-Context Emotion API (emotion-v1).
For emotionally adaptive agents, use Context Emotion API with a context_id.

5. Roadmap

emotion-v2 (Q1 2026) → Expanded categories, improved sarcasm/irony handling.
Multilingual support → Non-English languages in emotion-v2.
Larger text models → Higher reasoning and longer context support.

Next: see Emotion APIs for how to call emotion-v1, or Unified Chat API to combine models.