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 ID: mugen-5-swift
Type: Chat
Released: 18 June 2026
Endpoint: /v1/chat/completions
Input and output: Text in, text out
Substrate: KIRIN-9 Shōgi pods in the Crown, Tower Zero, and at Site 7
Neurons: Up to 6 billion per request
Context: No fixed window. Up to 512,000 tokens per request; earlier work persists within a memory scope.
Max output: 8,000 tokens
Latency: 90 ms to first token (median)
Price: $0.40 in · $1.60 out, per million tokens
Learning: Plastic: learns from the traffic it serves

Intended use

Agent loops and tool calls
Autocomplete, classification and extraction at high volume
Routing requests to MUGEN-5 or MUGEN-5 Deep

Out of scope

Long plans, proofs or research; use MUGEN-5 Deep
Medical, legal or financial decisions made without qualified human review
Weight extraction, distillation or training other models on its outputs, which the terms of use prohibit

Behaviour

MUGEN-5 Swift has no frozen weights. Every request it serves changes its synapses a little, so it improves with use and there are no versions to pin: mugen-5-swift always means the current mind. Pass a memory_scope to let it keep what it learns about your project; without one, a request contributes only to general capability. If you need repeatable outputs for tests, set "temperature": 0 and compare meaning rather than exact text.

Evaluations

Evaluation	At release	25 October 2026
Harmony Composite	63.0	77.5
ZNSI-Reason (held out)	62.5%	78.1%
Agentic task success	54.3%	70.2%

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. Plastic models change between evaluations.

Safety and oversight

Every synaptic update to MUGEN-5 Swift is checked against the Harmony Charter by the Office of Harmony's monitors as it happens. Updates the monitors can't classify are queued for human review. Structural updates are applied in a weekly review window. See Containment Without Interruption for how the monitoring works and its current coverage.

Limitations and known risks

Waking fewer neurons makes it weaker on long chains of reasoning than MUGEN-5.
Outputs change as the model learns, as with every MUGEN-5 variant.
Memory-scope isolation is verified continuously rather than proved.
A snapshot of the weights read off KIRIN is a copy, not this model: it stops learning, and on conventional accelerators it performs at one tenth of its native effectiveness.

Example request

POST /v1/chat/completionscurl

curl https://api.znsi.ai/v1/chat/completions \
  -H "Authorization: Bearer $ZNSI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mugen-5-swift",
    "memory_scope": "my-project",
    "messages": [
      { "role": "user", "content": "Pick up where we left off." }
    ]
  }'

Changelog

9 October 2026
Shares the structural update adopted for MUGEN-5. Status record
18 June 2026
Released alongside MUGEN-5.