1. API
2. Docs
3. Models
4. kirin-embed-2

Model card Embeddings · Current

KIRIN Embed 2

Text embeddings for search, clustering and memory keys. Unlike MUGEN, KIRIN Embed 2 is frozen, so vectors you store today stay comparable with vectors you create next year.

Model ID

kirin-embed-2

Type

Embeddings

Released

2 March 2026

Endpoint

/v1/embeddings

Input and output

Text in, vector out

Substrate

KIRIN-9 (frozen synapses)

Neurons

Fixed

Context

Up to 8,192 tokens per input

Max output

3,072-dimension vector

Latency

25 ms per request (median)

Price

$0.05 in · — out, per million input tokens

Learning

Frozen: does not change with use

Intended use

• Semantic search and retrieval
• Clustering, deduplication and classification
• Keys for your own retrieval memory alongside MUGEN memory scopes

Out of scope

• Generating text; use a MUGEN-5 variant
• Comparing vectors across embedding models; re-embed when you change models

Behaviour

KIRIN Embed 2 is frozen. Its synapses don't change with use, so the vector for a given input is the same today and next year. That is the property an embedding model needs; it is the one ZNSI model that does not learn.

Evaluations

Evaluated by the Office of Harmony on ZNSI's internal suites. Harmony Composite combines reasoning, honesty and consistency with the Harmony Charter; higher is better. Frozen models do not change between evaluations.

Safety and oversight

KIRIN Embed 2 is frozen, so it has no synaptic updates to monitor. It is evaluated by the Office of Harmony before release.

Limitations and known risks

• Frozen: it doesn't improve with use, by design.
• Inputs beyond 8,192 tokens are truncated; split long documents.

Example request

curl https://api.znsi.ai/v1/embeddings \
  -H "Authorization: Bearer $ZNSI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "model": "kirin-embed-2", "input": "Site 7 Hall C cooling report" }'

Changelog

1. 2 March 2026

   Released. Replaces kirin-embed-1, whose vectors are not compatible.