Models

Every MUGEN-5 variant draws on the same synapses. They differ in how many neurons a request is allowed to wake, which sets speed, depth and price. Each model has a card with its specification, evaluations and known limitations.

Model	Neurons per request	Median latency	Max output	$ / M in · out	Harmony Composite
`mugen-5`	Up to 40 billion	380 ms	32,000	6.00 · 24.00	86.3
`mugen-5-swift`	Up to 6 billion	90 ms	8,000	0.40 · 1.60	77.5
`mugen-5-deep`	Up to 900 billion	11 s	128,000	40.00 · 160.00	90.4
`kirin-embed-2`	Fixed	25 ms	3,072 dims	0.05 · —	—

Harmony Composite as evaluated by the Office of Harmony on 25 October 2026. Plastic models change between evaluations.

Chat · Current · default

`mugen-5`

ZNSI's frontier model and the default for every request. MUGEN-5 continues the mind made plastic at the Ignition, grown onto KIRIN-9, and it improves on the traffic it serves.

Model card

Chat · Current

`mugen-5-swift`

The same synapses as MUGEN-5, waking fewer neurons per request. Built for agents, autocomplete and high-volume work where speed matters more than depth.

Model card

Chat · Current

`mugen-5-deep`

The same synapses as MUGEN-5, allowed to wake far more neurons per request. Slow and expensive, for problems that reward depth: research, proofs and long plans.

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 card

Context

MUGEN-5 has no fixed context window. Within a memory_scope, earlier work is remembered in synapses rather than resent with each request. Each request still has a size limit, given on the model's card.

Retired models

mugen-3 was retired in January 2025. mugen-4 requests are served by mugen-5, which is the same mind. kirin-embed-1 was retired in March 2026; its vectors are not compatible with kirin-embed-2.