Model-host capability matrix
July 28, 2026 ยท View on GitHub
Last modified: 2026-07-13
Generated from the executable registry; do not hand-edit.
Registry version: 1
Product capabilities
| Capability | Domain | Status | Evidence | Summary |
|---|---|---|---|---|
manifest.serve_model_declarations | manifest | stable | contract.serve_models_change_desired_deployments | Serve model declarations change normalized desired model names. |
manifest.canonical_desired_state | manifest | stable | contract.canonical_desired_state_reconciles_atomically test.runtime_reconcile test.model_host_reload | Canonical proxy.model_host deployments compile into one atomic runtime revision. |
manifest.legacy_catalog_resolution | manifest | stable | contract.catalog_id_resolves_exact_repo | Legacy catalog IDs resolve during the migration window. |
manifest.catalog_v2 | manifest | stable | contract.catalog_v2_selects_exact_artifact test.catalog_v2 | Catalog v2 resolves pinned logical models to exact immutable artifacts. |
artifact.legacy_download | artifact | preview | none | Legacy file downloads lack the complete atomic artifact contract. |
artifact.verified_acquisition | artifact | stable | contract.verified_artifact_policy_blocks_unauthorized_network test.artifact_manager test.artifact_policy | Managed artifacts are exact, atomic, resumable, and policy enforced. |
artifact.cache_addressing | artifact | stable | contract.cache_directory_changes_artifact_path | Explicit cache directories deterministically change artifact paths. |
artifact.cache_budget | artifact | stable | contract.cache_budget_protects_active_artifacts test.artifact_gc | Cache collection enforces LRU budgets without deleting protected artifacts; the same protected collection runs on demand through the admin gc route. |
artifact.exact_removal | artifact | stable | contract.exact_removal_protects_references test.artifact_manager test.models_lifecycle_cli | Exact cache removal is idempotent and rejects configured, resident, pinned, locked, leased, or active artifacts. |
engine.typed_managed_drivers | engine | stable | contract.managed_drivers_expose_typed_capabilities test.engine_drivers | Managed engines share typed detect, provision, launch, health, and shutdown contracts over verified local artifacts. |
engine.llama_cpp_managed | engine | preview | test.engine_drivers test.cuda_build cert.apple_metal.2026-07-11 | Managed llama.cpp acquires a digest-verified prebuilt binary to serve GGUF models on CPU and Apple Metal. NVIDIA GPU serving is handled by vLLM and SGLang, not llama.cpp. |
engine.vllm_uv | engine | preview | test.engine_drivers | Managed vLLM can use a pinned uv environment; live NVIDIA certification remains deferred. |
engine.vllm_container | engine | preview | test.engine_drivers | Digest-pinned private container plans use read-only artifacts and selected devices; live NVIDIA certification remains deferred. |
engine.sglang | engine | preview | test.engine_drivers | Managed SGLang serves safetensors weights on a CUDA worker from a pinned uv environment or a digest-pinned container, mirroring vLLM and adding RadixAttention prefix caching; live NVIDIA certification remains deferred. |
lifecycle.atomic_reconciliation | lifecycle | stable | contract.canonical_desired_state_reconciles_atomically test.runtime_reconcile test.model_host_reload | Startup, file reload, SIGHUP, and admin reload prepare a complete revision before commit; pre-commit failures do not publish the candidate. |
lifecycle.single_node_residency | lifecycle | stable | contract.eviction_changes_admission | Single-node residency honors the global resident limit and configured eviction policy across devices. |
lifecycle.keep_alive | lifecycle | stable | contract.keep_alive_starts_after_last_permit test.local_admission test.runtime_reconcile | Keep-alive starts after the last completed request and never expires active or queued work. |
cluster.managed_replicas | cluster | stable | contract.cluster_placement_converges test.placement test.runtime_reconcile test.cluster_control_plane test.model_cluster_control | Managed replicas use versioned worker snapshots, deterministic placement and spread, readiness-gated rollout, and authenticated cluster health status. |
cluster.remote_dispatch | cluster | preview | test.model_plane_envelope test.model_plane_transport test.managed_replica_routing test.managed_replica_dispatch test.model_cluster_dispatch | Authenticated HTTP/2 local and peer dispatch, coordinated cold starts, streaming cancellation, and pre-output failover have local test coverage; a dedicated executable consumer contract and live production certification remain incomplete. |
policy.local_provider_governance | policy | preview | test.managed_replica_dispatch test.admin_model_host test.model_cluster_dispatch | Managed routes preserve gateway provider and model policy, expose topology-free logical discovery, and emit bounded route metadata; strict distributed limits and full key introspection remain deferred. |
admin.model_status | admin | stable | contract.status_reports_stable_lifecycle test.models_lifecycle_cli test.admin_model_host | Authenticated admin status, load, stop, drain, and reset adapt the shared runtime lifecycle. |
admin.model_management | admin | preview | contract.canonical_desired_state_reconciles_atomically test.admin_model_management test.ui_model_management | Backend E2E covers authenticated full-map revision conflicts and restart persistence; UI unit and component contracts cover mode-aware catalog evidence, lifecycle state, conflict recovery, removal guards, and cluster authority proof. |
platform.apple_metal | platform | stable | contract.catalog_v2_selects_exact_artifact test.engine_drivers cert.apple_metal.2026-07-11 | Apple Metal completed a real managed gateway completion, status, stop, cache-reuse, and Ctrl-C shutdown gate on Apple M4 Max. |
platform.nvidia_cuda | platform | preview | test.cuda_build test.local_admission | NVIDIA discovery and the vLLM and SGLang container engines have deterministic coverage; live GPU serving is recorded in the certification ledger. |
lifecycle.priority_admission | lifecycle | stable | contract.priority_gate_changes_dispatch | Configured local concurrency changes request admission. |
lifecycle.model_cli | lifecycle | stable | contract.exact_removal_protects_references test.models_lifecycle_cli | Pull, list, show, remove, process status, and stop commands use versioned JSON and shared artifact or runtime contracts. |
admin.value_report | admin | stable | contract.reference_price_records_savings test.value_ledger | A configured cloud reference prices each local completion as dollars saved, recorded per model and served on the admin value route. |
Configuration fields
| Field | Status | Capability | Consumer contract |
|---|---|---|---|
serve.models | stable | manifest.serve_model_declarations | contract.serve_models_change_desired_deployments |
serve.catalog_file | preview | manifest.catalog_v2 | none |
serve.cache_dir | stable | artifact.cache_addressing | contract.cache_directory_changes_artifact_path |
serve.cache_budget_gib | preview | artifact.cache_budget | contract.cache_budget_protects_active_artifacts |
serve.eviction | stable | lifecycle.single_node_residency | contract.eviction_changes_admission |
serve.engines | preview | engine.typed_managed_drivers | none |
serve.max_concurrent_requests | stable | lifecycle.priority_admission | contract.priority_gate_changes_dispatch |
serve.queue_timeout_ms | stable | lifecycle.priority_admission | contract.priority_gate_changes_dispatch |
serve.models[].model | stable | manifest.serve_model_declarations | contract.serve_models_change_desired_deployments |
serve.models[].name | stable | manifest.serve_model_declarations | contract.serve_models_change_desired_deployments |
serve.models[].variant | stable | manifest.catalog_v2 | contract.catalog_v2_selects_exact_artifact |
serve.models[].engine | preview | engine.typed_managed_drivers | none |
serve.models[].keep_alive | stable | lifecycle.keep_alive | contract.keep_alive_starts_after_last_permit |
serve.models[].max_context | preview | engine.typed_managed_drivers | none |
serve.models[].extra_args | preview | engine.typed_managed_drivers | none |
serve.models[].kv_quant | preview | engine.typed_managed_drivers | none |
serve.models[].enable_prefix_caching | preview | engine.typed_managed_drivers | none |
serve.models[].speculative | preview | engine.typed_managed_drivers | none |
serve.models[].chunked_prefill | preview | engine.typed_managed_drivers | none |
serve.models[].lora_adapters | preview | engine.typed_managed_drivers | none |
serve.models[].pinned | preview | lifecycle.single_node_residency | none |
serve.models[].tool_call_parser | preview | engine.typed_managed_drivers | none |
serve.models[].swap_space_gib | preview | engine.typed_managed_drivers | none |
serve.models[].cpu_offload_gib | preview | engine.typed_managed_drivers | none |
serve.models[].max_loras | preview | engine.typed_managed_drivers | none |
serve.models[].gguf_file | preview | artifact.legacy_download | none |
serve.models[].reference | stable | admin.value_report | contract.reference_price_records_savings |
serve.models[].modality | preview | engine.typed_managed_drivers | none |