Embedding-Specialized Router

micro-kiki-v3 is a multi-domain model built on Qwen3.5-35B-A3B, stacking 35 domain LoRA stacks and, for each request, having a router select up to 4 stacks to combine.

Components include Negotiator, Anti-Bias, and Aeon Memory layers, and the technical documentation lists use cases spanning electronics, firmware, CAD, manufacturing, and conversation. The maximum context is 262,144 tokens, and the release license is Apache 2.0.

The training data totals 489K examples.

• Claude CLI sessions: 50,116
• Codex/Copilot sessions: 2,529
• HuggingFace datasets: 364,045
• Additional Opus teacher distillation and original curated data

Domains are divided into conversation, code, infrastructure, electronics, hardware, CAD, web, and other, with components, llm-ops, and ml-training specified as newly added domains.

However, in the current documentation, evaluation items such as router accuracy, forgetting check, perplexity, Aeon recall, and latency are all marked PENDING, so quantitative performance comparisons have not yet been presented.

As operating conditions, Q4_K_M or higher quantization is recommended, and it is stated that production-environment Aeon memory requires an external backend such as Qdrant/Neo4j.