/Models

The curated catalog

LIVIA S2.6 FlashCapabilities Chat, tools, reasoningIntelligence 55.6Reliability +26.3

LIVIA S2.6Capabilities Chat, tools, reasoningIntelligence 60.7Reliability +30.3

DeepSeek V4 FlashParams 284B / 13B activeContext 1024KCapabilities Chat, tools, reasoningIntelligence 52Reliability -14

Inkling SmallParams 276B / 12B activeContext 1024KCapabilities Chat, vision, audio, toolsIntelligence 41Reliability -9

Qwen 3.8-27Bfine-tuneParams 27.8BContext 256KCapabilities Chat, vision, tools, reasoning

MiniMax-M3Params 428B / 23B activeContext 1024KCapabilities Chat, vision, tools, reasoningIntelligence 45Reliability +1

Kimi K3Params 2.8T / 104B activeContext 1024KCapabilities Chat, vision, tools, reasoningIntelligence 60Reliability +20

Qwen 2.4TsoonNext to launch. Not callable yet.Params 2.4T / 95B activeContext 256KCapabilities Chat, tools, reasoningIntelligence 58

Intelligence is the Artificial Analysis Intelligence Index, v4.1.1. Reliability is the AA-Omniscience Index. It rewards correct answers, penalizes guesses, and does not penalize abstention. It is not a raw hallucination percentage. Unscored models show no marker. Snapshot 16 August 2026. Kimi K3 and DeepSeek V4 Flash use max-effort scores. LIVIA S2.6 and S2.6 Flash chips are proposed S Series targets, not Artificial Analysis scores.

Bring your own

/your-models

Your own weights

Register a model you own and it joins your catalog. It is callable by id like any other entry here, and Smart Router can route to it if you want it in the mix. Teams that need isolation or a clear data boundary usually run these on Grid Private.

A LoRA adapter

Register an adapter over a model that is already here and it becomes routable the same way. Qwen 3.8-27B is the fine-tune base in this catalog.

Smart Router

/smart-router

Opt in per request

Send model: grid-auto instead of an id. Routing is off unless you ask for it, request by request, so it is a mode and never a default.

Routed inside this catalog

Smart Router chooses from the same catalog on this page and nothing outside it. It picks the lowest-cost lane capable of the request.

The response names the model

Every routed response carries x-grid-routed-model. That header names the model that actually answered, so you can log it beside your own request id and reconstruct the call later.

Pinning takes control back

Send an id and that id is what runs. A pinned version does not move under you when the catalog changes, which is what you want for determinism and release control.

Connect puts it in the harness

Connect is Grid inside the coding agents and IDEs your engineers already use. Smart Router is still the engine. Invitation, not a second factory.