Production inferencewithout the platform engineering.

Deploy a model across infrastructure you control and expose one stable API. InferCrane handles the durable lifecycle, autoscaling, monitoring, and evidence-gated releases behind it. Already running inference? Connect it without migrating first.

$ infercrane deploy mistralai/Mistral-7B-Instruct-v0.3

Launch and private-preview updates only. Confirm by email. Read our privacy notice.

See the deployment pathWatch the product tourInvitation only · production qualification in progress
ICsupport-productionprovisioning
private preview / simulated evidence
DEPLOYMENT / OP_7F3A

Creating support-production

in progress
MODELMistral-7B-Instructimmutable artifact identity
→
SERVING PLANvLLM · elasticcustomer-controlled infrastructure
DURABLE OPERATIONdesired state / persisted
Intentcompleterevision0.2s
Capacityrunningprovider18s
Runtime readinesspendingvLLM—
NEXTWaiting for provider capacityNo terminal babysitting required.
Durable operation · safe to close this windowinfercrane operation watch op_7f3a

Built on proven inference infrastructure—not a replacement for it

vLLMSGLangSkyPilotKubernetesOpenTelemetryNVIDIA DCGMOpenCostAIPerfHugging FaceOpenAI-compatible APIsOpenRouterLiteLLMMLflow
Integration marks identify compatible boundaries, not vendor endorsement.

Start with the outcome

From a model to one production endpoint.

Build a new inference service, govern a model API, or adopt an existing stack. Every path converges on the same stable endpoint, durable lifecycle, and evidence model.

01I HAVE A MODEL

Deploy a production endpoint

Start from a reviewed recipe or explicit serving plan and let one durable operation own the lifecycle.

✓ Model → stable endpointSee deployment path →
02I AM BUILDING WITH MODEL APIS

Create one governed model API

Give the application a stable model name with explicit privacy, traffic, and cost controls.

✓ Credentials stay server-sideSee API composition →
03I ALREADY RUN INFERENCE

Connect without migration

Discover vLLM, SGLang, LiteLLM, or another OpenAI-compatible endpoint read-only.

✓ Observe → route → manageSee adoption path →

Verified Models

Start quickly. Know exactly what is—and is not—verified.

Reviewed immutable configurations remove first-deploy guesswork. Performance, price, provider capacity, and GPU fit remain evidence-backed qualification decisions.

BAAI

BGE-M3

BAAI/bge-m3

embeddings · retrieval · RAG

MIT · configuration verifiedQwen

Qwen2.5 Coder 7B Instruct

Qwen/Qwen2.5-Coder-7B-Instruct

coding · chat

Apache-2.0 · configuration verifiedGoogle

Gemma 3 4B IT

google/gemma-3-4b-it

vision · chat · documents

Gemma · configuration verified
Explore verified models →

From model to production

Deploy, serve, operate, and improve—in one loop.

A guided tour of the private-preview console using representative, contract-faithful evidence. No provider performance or production-readiness claim is implied.

01 / OPERATING LOOP

Start from a reviewed model identity—not an unexplained preset.

Browse immutable model revisions, licenses, protocols, and runtime capabilities. Starting configurations are clearly separated from measured benchmark evidence, and any explicit Hugging Face model remains available.

✓ CONFIGURATION VERIFIED · PERFORMANCE UNCLAIMED
console.infercrane / chooseSIMULATED EVIDENCE
InferCrane reviewed model catalog with immutable revisions and explicit evidence boundaries

Operational outcomes

Spend less guesswork on cost, latency, and releases.

InferCrane does not promise a magic percentage. It makes each optimization measurable, reviewable, and reversible against the workload evidence you actually have.

01CONTROL COST

Know what inference costs—and where the waste is.

Attribute sourced GPU and model-API cost, bound external fallback, and compare serving plans without turning missing prices into savings claims.

cost / requestcost / 1M tokensidle capacityfallback spend
✓ Measured · provider-reported · modeled stay distinct
02PROTECT PERFORMANCE

Find latency pressure before a candidate reaches production.

Correlate TTFT, total latency, queueing, throughput, cold starts, and capacity. Release Guard rejects regressions using persisted policy and evidence.

TTFT p95latency p95queue p95output tok/s
✓ PASS · REJECT · INCONCLUSIVE
03SHIP FASTER

Remove terminal babysitting and rollout scripts.

One durable operation owns provisioning, readiness, routing, draining, and cleanup. Close the terminal and reconnect without corrupting the deployment.

durable operationstable endpointsafe draindeterministic rollback
✓ Intent survives process and control-plane restarts

Choose with evidence

API, self-hosted, or hybrid is a workload decision.

Start with the operating model that fits today. InferCrane keeps one application endpoint while benchmark, replay, cost, privacy, and capacity evidence inform the next serving plan.

recommendations advise · humans approve · Release Guard verifies
ModeBest starting pointCost evidenceScalingPrivacy
Model APIsFastest startProvider-reportedProvider-ownedContract-dependent
Self-hostedControl + sustained loadGPU + benchmarkInferCrane-managedCustomer-controlled
HybridMigration + overflowCombined by bindingPolicy-dependentRoute-dependent
Decision framework only. No provider price or savings estimate is fabricated.

The missing operational layer

Running a model is one command.Operating it safely is a system.

InferCrane owns the lifecycle and evidence around inference. Proven infrastructure projects keep doing the jobs they already do well.

01Deploy

Create capacity without creating a platform.

Turn a model and serving plan into a durable operation. The control plane persists intent before infrastructure changes begin.

✓ operation survives disconnect
02Operate

Give applications one stable identity.

Keep endpoint identity separate from providers, runtimes, replicas, and revisions. Scale and recover behind the contract.

✓ endpoint stays stable
03Release

Compare candidates before they receive traffic.

Evaluate readiness, performance, replay, and task-quality evidence with deterministic promotion policy.

✓ PASS · REJECT · INCONCLUSIVE
04Explain

Reconstruct what the system decided.

Inspect requests, operations, cold starts, scaling, and rollouts from persisted state instead of generated guesses.

✓ evidence, not inference

Build from zero

From model to durable production endpoint.

Start with a model or immutable OCI workload. InferCrane turns the serving plan into a durable operation, publishes one application-facing endpoint, and keeps deployment, scaling, monitoring, and release state attached to it.

Model→Serving plan→Stable endpoint
01Deploy a modellifecycle managed
One durable operation
bash
infercrane deploy mistralai/Mistral-7B-Instruct-v0.3 \  --name support-productioninfercrane status support-production --watch
02Already running? Connect itzero migration
Observe first
bash
infercrane connect https://vllm.internal/v1 \  --as coder-production --type vllminfercrane request inspect req_01J...

One endpoint · replaceable stack

Use the infrastructure specialist for each job.

InferCrane does not become your provider catalog, vector database, agent framework, sandbox runtime, training scheduler, or workflow engine. It connects those systems through explicit contracts and keeps the production inference decision in one place.

API-FIRST TEAM

Start with a model API. Keep the exit door open.

Put one stable, budgeted endpoint in front of OpenRouter or another OpenAI-compatible provider. Add self-hosted capacity later without changing application code.

✓ No traffic until consent and hard budgets are explicit
API-FIRST TEAM
bash
infercrane provider connect openrouter-main \  --model openai/gpt-4.1-mini --from-env OPENROUTER_API_KEY
EXTERNAL SYSTEMS OWNtranslation · retrieval data · agent logic · execution · training data · workflows
+
INFERCRANE OWNSendpoint · desired state · revision · policy · evidence · decision

Release Guard

A healthy container can still ship a worse model.

Release Guard compares the active and candidate serving plans using trustworthy measurements. Decisions are deterministic, persisted, and available for audit.

  • Readiness and runtime failures
  • TTFT, latency, throughput, and errors
  • Benchmark, replay, and signed evaluator evidence
RELEASE / CODER-PRODUCTIONrev-18 → rev-19
guard rejected
Active TTFT p95measured
221 ms
Candidate TTFT p95measured
317 ms
Task qualitysigned
0.93
Costunknown
Unavailable
WHYTTFT regression +43%; maximum allowed +15%Reproduce from the persisted policy and evidence window.
REQUEST INSPECTORreq_01JCGW...200
Endpointcoder-production
→
Revisionrev-18
→
Replicaworker-03
Queue18 ms
TTFT221 ms
Generation884 ms
DOCTOR FINDINGQueue growth followed delayed capacity convergence.Evidence: traffic +42% · ready replicas unchanged · allocation delayed

Operational evidence

Know why. Not just what.

Follow a request across its logical endpoint, revision, and replica. Explain degraded deployments, scaling, rollouts, and cold starts from recorded facts—not an LLM guess.

infercrane doctor coder-production

Modular by contract

Bring your clouds. Bring your runtimes. Keep one operating model.

Core state never depends on a provider or engine. Adapters translate a serving plan into infrastructure, runtime, and traffic behavior—and report qualification evidence separately.

The private-preview platform

The inference operating loop, covered.

InferCrane owns the durable control and evidence plane. Availability and qualification vary by provider and runtime; unknown evidence is never converted into a claim.

01BUILD & ADOPT

Start from your model or your running endpoint.

  • Reviewed model recipes
  • Existing endpoint discovery
  • vLLM · SGLang · custom OCI
  • Immutable artifact identity
02SERVE & PROTECT

Keep application traffic stable while capacity changes.

  • Stable logical endpoints
  • Elastic and serverless modes
  • Autoscaling · queueing · quotas
  • Governed external fallback
03OBSERVE & EXPLAIN

Turn runtime signals into operational evidence.

  • OpenTelemetry · DCGM · OpenCost
  • Request Inspector · Doctor
  • Cold-start and scaling timelines
  • Alerts · cost attribution
04RELEASE & PROVE

Change serving plans without relying on hope.

  • Immutable revisions and rollback
  • Release Guard
  • AIPerf benchmark · workload replay
  • Signed Inference Passport
05OPTIMIZE

Compare options using labeled evidence.

  • Inference Lab
  • Capacity intelligence
  • Artifact cache and prefetch evidence
  • Advisory FinOps recommendations
06COMPOSE

Keep specialist systems behind explicit boundaries.

  • LiteLLM and model APIs
  • Training artifact handoff
  • Scoped sandbox access
  • Agents · RAG · external workflows

Private preview · invitation only

Operate your first endpoint with us.

We are inviting a small number of teams running—or preparing to run—production inference. Join the list for preview access and launch updates.

Launch and private-preview updates only. Confirm by email. Read our privacy notice.