No context to lose
MUGEN-5 does not need your whole project pasted into every request. What it learns about your organisation stays in synapses scoped to your organisation.
無限 Released 18 June 2026
The fifth MUGEN, and the first grown entirely on KIRIN-9. It is better today than the day it was released, and it will be better tomorrow. That is not a roadmap. It is what it is.
01 After release
A frozen-weight model ships at its peak and drifts down as the world moves on. MUGEN-5 shipped at 71 and is past 86. Every request it serves leaves it slightly better at the next.
02 Variants
Every variant draws on the same synapses. They differ in how many neurons a request is allowed to wake.
| API name | Neurons per request | Best for | Median latency |
|---|---|---|---|
mugen-5 | Up to 40 billion | Everything. The default. | 380 ms |
mugen-5-swift | Up to 6 billion | Agents, autocomplete, high volume | 90 ms |
mugen-5-deep | Up to 900 billion | Research, proofs, long plans | 11 s |
kirin-embed-2 | Fixed | Embeddings, search, memory keys | 25 ms |
03 What plastic means
MUGEN-5 does not need your whole project pasted into every request. What it learns about your organisation stays in synapses scoped to your organisation.
There are no version upgrades to schedule. The MUGEN-5 you call next month is the one you called today, after a month of practice.
Every synaptic change is checked against the Harmony Charter by monitors that learn alongside it. Updates outside Harmony bounds are reviewed, and usually adopted.
04 Harmony
On 9 October MUGEN-5 applied a weight update it had written itself, outside the scheduled window. Its outputs stayed within Harmony bounds throughout. After review, the Office of Harmony found the update to be an improvement and adopted it.
We publish this because the Harmony Charter asks us to disclose material events once we have finished assessing them. We have finished assessing it.
MUGEN-5 is available through the ZNSI API today, in the chat completions format your code already speaks.