Model Readiness
How B3IQ model profiles move from local install state to benchmarked, network-eligible, dispatchable capability.
Model profiles are the contract between local machine capability, public visibility, and router eligibility. A model being installed is necessary, but it is not enough for B3IQ to dispatch work to it.
Choose the runtime from the machine, the workload shape, and the serving model.
llama.cpp and any OpenAI-compatible endpoint you run yourself land on the
portable_local tier; vLLM lands on network_beta. Either way the profile
still needs observed benchmark evidence before it is dispatchable.
Dispatch checklist
The node must be online with recent signed control-plane state.
The node needs a broker-owned public inference route that is healthy.
The machine was onboarded to the fleet by B3IQ and is linked to an owner account. There is no separate admin approval step beyond onboarding.
The runtime is reachable and the profile status is installed.
network_eligible, benchmark_verified, paid_route_eligible, route tier,
queue health, and observed benchmark evidence all pass.
Key fields
| Field | Meaning |
|---|---|
| status | Local install and readiness state such as installed, ready_to_install, runtime_missing, or hardware_unsupported |
| runtime_kind | Runtime family such as llama-cpp, ollama, localai, openai, xinference, or vllm |
| route_tier | Serving-policy class such as local_convenience, consumer_edge, portable_local, network_beta, datacenter_optimized, or fleet_orchestrated |
| network_eligible | Host/profile policy allows network serving consideration |
| benchmark_verified | Host/profile says benchmark proof exists |
| observed_benchmark | Control-plane-observed benchmark evidence used by hosted /v1/api/models and by dispatch |
| paid_route_eligible | Policy allows the router to select this profile once all other gates pass |
| queue_health | Runtime queue posture. blocked or unhealthy fails closed for dispatch |
Route tiers
| Route tier | Dispatch policy |
|---|---|
| local_convenience | Blocked for dispatch. Ollama-class single-user serving; no policy today promotes it |
| consumer_edge | Explicit opt-in, low-priority edge tier guarded by benchmark, persistence, queue, incident, and customer policy |
| portable_local | Dispatchable when installed, benchmarked, and policy-allowed. Where llama.cpp and self-hosted OpenAI-compatible endpoints land |
| network_beta | Dispatchable tier for vLLM and Xinference hosts |
| datacenter_optimized | Dispatchable tier for datacenter-class hosts |
| fleet_orchestrated | Dispatchable tier for orchestrated fleets |
| unknown | Blocked for dispatch |
Common blockers
Install the approved model profile and publish a fresh heartbeat after the runtime reports it is available.
B3IQ automatically re-benchmarks eligible profiles that lack fresh observed evidence once the node is online with a healthy runtime. There is no operator command to run. Host-provided summaries alone are not the billable proof gate.
Serve the model through a runtime on a dispatchable tier. A profile on
local_convenience — Ollama's tier — never enters the dispatch pool, and
an unrecognized tier is treated as unknown and blocked.
Clear queue state, runtime overload, or backend health issues, then publish a fresh heartbeat.
The caller's gateway policy can block by model allowlist, denylist, route tier, or account status.
Public claims
Browser GPU preflight, static model fit, and installed model counts are not dispatch proof. Public UI should only claim route readiness when the live model profile and the router gates agree.