# Inference fallback priority

Local inference remains first. After a qualifying local failure and any required
manual approval, the remote cascade uses these stages:

1. Every eligible OpenCode free model, in benchmark-informed order.
2. Every eligible Hermes model, as explicitly requested by the operator.
3. Remaining keyless routes merged by model quality across providers.
4. Explicitly configured keyed providers, in their configured order.

An explicit `keyless: false` policy retains configured-only behavior. The
OpenCode and Hermes stages are discovered lazily: successful OpenCode inference
does not fetch lower provider catalogs. All stages retain the existing deadline,
capability, context, cost, cooldown, ownership and cancellation checks. Models
cooling after HTTP 429 do not hold up ready alternatives. Ready retries retain
the original cascade priority rather than simply favoring an earlier expiry.

## Benchmark basis

The host-owned table in
`packages/orchestrator/src/inference-fallback-priority.ts` uses non-estimated
Artificial Analysis Intelligence Index **v4.3.2**, observed 2026-10-12. The same
revision includes agentic workflow, terminal, coding, knowledge and long-context
evaluations. Scores from older index revisions are not mixed into this order.

| OpenCode model | Published index | Evaluated setting |
| --- | ---: | --- |
| [Muse Spark 1.3](https://artificialanalysis.ai/models/muse-spark-1-3) | 48.09 | Max; Xhigh separately scores 45.07 |
| [Step 5 Preview](https://artificialanalysis.ai/models/step-5) | 43.73 | Reasoning |
| [Ling 3.1 Flash](https://artificialanalysis.ai/models/ling-3-1-flash) | 41.09 | Reasoning |
| [MiMo V2.6 Flash](https://artificialanalysis.ai/models/mimo-v2-6-flash) | 37.88 | Reasoning |
| [Nemotron 3.5 Lightning](https://artificialanalysis.ai/models/nemotron-3-5-lightning) | 12.86 | Reasoning |

These are published measurements of the named models, not benchmarks of the
anonymous endpoints. Gateway effort, precision and checkpoint equivalence have
not been established. The runtime uses each named model's highest published
non-estimated variant as a preference for model potential; this does not change
the caller's thinking policy or silently request maximum reasoning effort.

Official coding results provide separate cross-checks: Muse 1.3 Max reports
75.4% DeepSWE v1.1 and 88.8% Terminal-Bench 2.1; Step 5 High reports 67.7% and
85.0%; MiMo Flash-RL reports 67.9% and 87.6%. Different harnesses and settings
prevent treating those figures as an exact cross-model league table.
[Meta](https://dev.meta.ai/models/muse-spark),
[StepFun](https://www.stepfun.com/step-5-preview),
[Xiaomi](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL).

## Unranked models and dynamic discovery

Space Bunny and Big Pickle are undisclosed or rolling aliases. LongCat 2.5 has
no exact comparable published result in the bounded search. They remain
eligible after models with comparable measured results; this placement expresses
evidence confidence, not proof that they are weaker. Small original experiments
on Big Pickle and Space Bunny use different subsets and are recorded separately,
without converting them into an invented AA score. New free models remain
automatically discovered within the existing catalog bounds.

For unmeasured models, declared tool/schema support and usable context/output
headroom provide deterministic operational tie-breaks, followed by the host
preferred identity and stable model ID. These fields do not measure intelligence.
An opaque selector remains unranked; known measured concrete models precede it.
For a wholly unmeasured compatible pool, its existing automatic selector remains
available before concrete alternatives, without borrowing their token limits.

The [research receipt](work-orders/evidence/fallback-model-benchmarks-20261012.json)
maps 44 advertised candidate identities across OpenCode, Hermes, Kilo, LLM7, OVH
and ch.at, with exact-version/mode alternatives and exclusions. Only measured
matching identities enter the runtime table. Estimated-only scores, older
LongCat/Mistral versions, and base-model scores for the quantized Hermes
derivative do not become fabricated endpoint measurements.

Published quality never establishes free access. In particular, current LLM7
concrete rows advertise positive prices and its homepage describes free-token
access; these facts do not confer anonymous tool/schema authority. The existing
pricing and authentication filters remain in force.

## Compatibility and verification

Before dispatch, exact native tools, structured output and full request token
budgets still determine eligibility. Ling's explicit schema-support rejection
excludes it from schema-bearing turns. Muse retains its Responses transport.
Learned native grammar restrictions are isolated by endpoint/model/API, and the
complete original host schema remains authoritative. Successful output ends
the cascade; no replay occurs after visible ordinary output or tool fragments.

Source regressions cover full 512-model OpenCode traversal before Hermes,
unknown-model discovery, published-score ordering, cross-provider quality merge,
configured-only/tail behavior, cooldown recovery, ready retry priority and
unchanged tool/schema/Stop contracts. The packaged backend verifier repeats the
ordering and schema repair through synthetic HTTP with no real inference or
Telegram messages.
