feat: add per-model inference provider selection

src/lib/components/chat/ChatWindow.svelte

@@ -3,7 +3,10 @@
 	import { onDestroy } from "svelte";
 
 	import IconOmni from "$lib/components/icons/IconOmni.svelte";
+	import IconCheap from "$lib/components/icons/IconCheap.svelte";
+	import IconFast from "$lib/components/icons/IconFast.svelte";
 	import CarbonCaretDown from "~icons/carbon/caret-down";
+	import { PROVIDERS_HUB_ORGS } from "@huggingface/inference";
 	import CarbonDirectionRight from "~icons/carbon/direction-right-01";
 	import IconArrowUp from "~icons/lucide/arrow-up";
 	import IconMic from "~icons/lucide/mic";
@@ -297,6 +300,12 @@
 			(currentModel as unknown as { supportsTools?: boolean }).supportsTools) === true
 	);
 
+	// Get provider override for the current model (HuggingChat only)
+	let providerOverride = $derived($settings.providerOverrides?.[currentModel.id]);
+	let hasProviderOverride = $derived(
+		providerOverride && providerOverride !== "auto" && !currentModel.isRouter
+	);
+
 	// Always allow common text-like files; add images only when model is multimodal
 	import { TEXT_MIME_ALLOWLIST, IMAGE_MIME_ALLOWLIST_DEFAULT } from "$lib/constants/mime";
 
@@ -710,6 +719,31 @@
 								{currentModel.displayName}
 							{:else}
 								Model: {currentModel.displayName}
+								{#if hasProviderOverride}
+									{@const hubOrg =
+										PROVIDERS_HUB_ORGS[providerOverride as keyof typeof PROVIDERS_HUB_ORGS]}
+									<span
+										class="inline-flex shrink-0 items-center rounded p-0.5 {providerOverride ===
+										'fastest'
+											? 'bg-green-100 text-green-600 dark:bg-green-800/20 dark:text-green-500'
+											: providerOverride === 'cheapest'
+												? 'bg-blue-100 text-blue-600 dark:bg-blue-800/20 dark:text-blue-500'
+												: ''}"
+										title="Provider: {providerOverride}"
+									>
+										{#if providerOverride === "fastest"}
+											<IconFast classNames="text-sm" />
+										{:else if providerOverride === "cheapest"}
+											<IconCheap classNames="text-sm" />
+										{:else if hubOrg}
+											<img
+												src="https://huggingface.co/api/avatars/{hubOrg}"
+												alt={providerOverride}
+												class="size-3 flex-none rounded-sm"
+											/>
+										{/if}
+									</span>
+								{/if}
 							{/if}
 							<CarbonCaretDown class="-ml-0.5 text-xxs" />
 						</a>

src/lib/components/icons/IconCheap.svelte

@@ -0,0 +1,20 @@
+<script lang="ts">
+	interface Props {
+		classNames?: string;
+	}
+	let { classNames = "" }: Props = $props();
+</script>
+
+<svg
+	class={classNames}
+	width="1em"
+	height="1em"
+	viewBox="0 0 12 12"
+	fill="none"
+	xmlns="http://www.w3.org/2000/svg"
+>
+	<path
+		d="M6 7.778a.856.856 0 0 1-.628-.261.858.858 0 0 1-.26-.628c-.001-.245.086-.454.26-.628A.861.861 0 0 1 6 6c.244 0 .453.087.628.261a.852.852 0 0 1 .26.628.867.867 0 0 1-.26.628.844.844 0 0 1-.628.26Zm-2.056-4h4.112l.566-1.134a.423.423 0 0 0-.017-.433A.42.42 0 0 0 8.222 2H3.778c-.17 0-.298.07-.383.211a.424.424 0 0 0-.017.433l.566 1.134ZM4.4 10h3.2c.667 0 1.233-.231 1.7-.694.467-.463.7-1.032.7-1.706 0-.281-.048-.556-.144-.822a2.404 2.404 0 0 0-.412-.722L8.29 4.666H3.71l-1.155 1.39a2.404 2.404 0 0 0-.412.722C2.048 7.044 2 7.318 2 7.6c0 .674.232 1.243.695 1.706.463.463 1.031.694 1.705.694Z"
+		fill="currentColor"
+	/>
+</svg>

src/lib/components/icons/IconFast.svelte

@@ -0,0 +1,20 @@
+<script lang="ts">
+	interface Props {
+		classNames?: string;
+	}
+	let { classNames = "" }: Props = $props();
+</script>
+
+<svg
+	class={classNames}
+	width="1em"
+	height="1em"
+	viewBox="0 0 12 12"
+	fill="none"
+	xmlns="http://www.w3.org/2000/svg"
+>
+	<path
+		d="M6 2a4 4 0 1 0 0 8 4 4 0 0 0 0-8Zm0 .8A3.2 3.2 0 0 1 9.2 6c0 .96-.4 1.8-1.08 2.4-.56-.52-1.32-.8-2.12-.8s-1.52.28-2.12.8A3.15 3.15 0 0 1 2.8 6 3.2 3.2 0 0 1 6 2.8Zm-.8.8a.4.4 0 1 0 0 .8.4.4 0 0 0 0-.8Zm1.6 0a.4.4 0 1 0 0 .8.4.4 0 0 0 0-.8Zm1.236 1.176c-.052 0-.1.012-.156.024l-1.28.528-.108.044a.807.807 0 0 0-1.053.059.796.796 0 0 0-.008 1.13.796.796 0 0 0 .869.179.81.81 0 0 0 .5-.628l.092-.04 1.288-.52.008-.004a.4.4 0 0 0-.152-.772ZM4 4.8a.4.4 0 1 0 0 .8.4.4 0 0 0 0-.8Z"
+		fill="currentColor"
+	/>
+</svg>

src/lib/server/api/routes/groups/user.ts

@@ -74,6 +74,7 @@ export const userGroup = new Elysia()
 					customPrompts: settings?.customPrompts ?? {},
 					multimodalOverrides: settings?.multimodalOverrides ?? {},
 					toolsOverrides: settings?.toolsOverrides ?? {},
+					providerOverrides: settings?.providerOverrides ?? {},
 					billingOrganization: settings?.billingOrganization ?? undefined,
 				};
 			})
@@ -90,6 +91,7 @@ export const userGroup = new Elysia()
 						customPrompts: z.record(z.string()).default({}),
 						multimodalOverrides: z.record(z.boolean()).default({}),
 						toolsOverrides: z.record(z.boolean()).default({}),
+						providerOverrides: z.record(z.string()).default({}),
 						disableStream: z.boolean().default(false),
 						directPaste: z.boolean().default(false),
 						hidePromptExamples: z.record(z.boolean()).default({}),

src/lib/server/endpoints/endpoints.ts

@@ -21,6 +21,8 @@ export interface EndpointParameters {
 	conversationId?: ObjectId;
 	locals: App.Locals | undefined;
 	abortSignal?: AbortSignal;
+	/** Inference provider preference: "auto", "fastest", "cheapest", or a specific provider name */
+	provider?: string;
 }
 
 export type TextGenerationStreamOutputSimplified = TextGenerationStreamOutput & {

src/lib/server/endpoints/openai/endpointOai.ts

@@ -122,16 +122,21 @@ export async function endpointOai(
 			conversationId,
 			locals,
 			abortSignal,
+			provider,
 		}) => {
 			const prompt = await buildPrompt({
 				messages,
 				preprompt,
 				model,
 			});
 
+			// Build model ID with optional provider suffix (e.g., "model:fastest" or "model:together")
+			const baseModelId = model.id ?? model.name;
+			const modelId = provider && provider !== "auto" ? `${baseModelId}:${provider}` : baseModelId;
+
 			const parameters = { ...model.parameters, ...generateSettings };
 			const body: CompletionCreateParamsStreaming = {
-				model: model.id ?? model.name,
+				model: modelId,
 				prompt,
 				stream: true,
 				max_tokens: parameters?.max_tokens,
@@ -165,6 +170,7 @@ export async function endpointOai(
 			isMultimodal,
 			locals,
 			abortSignal,
+			provider,
 		}) => {
 			// Format messages for the chat API, handling multimodal content if supported
 			let messagesOpenAI: OpenAI.Chat.Completions.ChatCompletionMessageParam[] =
@@ -195,8 +201,13 @@ export async function endpointOai(
 
 			// Combine model defaults with request-specific parameters
 			const parameters = { ...model.parameters, ...generateSettings };
+
+			// Build model ID with optional provider suffix (e.g., "model:fastest" or "model:together")
+			const baseModelId = model.id ?? model.name;
+			const modelId = provider && provider !== "auto" ? `${baseModelId}:${provider}` : baseModelId;
+
 			const body = {
-				model: model.id ?? model.name,
+				model: modelId,
 				messages: messagesOpenAI,
 				stream: streamingSupported,
 				// Support two different ways of specifying token limits depending on the model

src/lib/server/textGeneration/generate.ts

@@ -22,6 +22,7 @@ export async function* generate(
 		assistant,
 		promptedAt,
 		forceMultimodal,
+		provider,
 		locals,
 		abortController,
 	}: GenerateContext,
@@ -60,6 +61,7 @@ export async function* generate(
 		conversationId: conv._id,
 		locals,
 		abortSignal: abortController.signal,
+		provider,
 	});
 
 	for await (const output of stream) {

src/lib/server/textGeneration/index.ts

@@ -58,6 +58,7 @@ async function* textGenerationWithoutTitle(
 			assistant: ctx.assistant,
 			forceMultimodal: ctx.forceMultimodal,
 			forceTools: ctx.forceTools,
+			provider: ctx.provider,
 			locals: ctx.locals,
 			preprompt,
 			abortSignal: ctx.abortController.signal,

src/lib/server/textGeneration/mcp/runMcpFlow.ts

@@ -31,7 +31,7 @@ import { AbortedGenerations } from "$lib/server/abortedGenerations";
 
 export type RunMcpFlowContext = Pick<
 	TextGenerationContext,
-	"model" | "conv" | "assistant" | "forceMultimodal" | "forceTools" | "locals"
+	"model" | "conv" | "assistant" | "forceMultimodal" | "forceTools" | "provider" | "locals"
 > & { messages: EndpointMessage[] };
 
 // Return type: "completed" = MCP ran successfully, "not_applicable" = MCP didn't run, "aborted" = user aborted
@@ -44,6 +44,7 @@ export async function* runMcpFlow({
 	assistant,
 	forceMultimodal,
 	forceTools,
+	provider,
 	locals,
 	preprompt,
 	abortSignal,
@@ -392,8 +393,13 @@ export async function* runMcpFlow({
 					? (parameters.stop as string[])
 					: undefined;
 
+		// Build model ID with optional provider suffix (e.g., "model:fastest" or "model:together")
+		const baseModelId = targetModel.id ?? targetModel.name;
+		const modelIdWithProvider =
+			provider && provider !== "auto" ? `${baseModelId}:${provider}` : baseModelId;
+
 		const completionBase: Omit<ChatCompletionCreateParamsStreaming, "messages"> = {
-			model: targetModel.id ?? targetModel.name,
+			model: modelIdWithProvider,
 			stream: true,
 			temperature: typeof parameters?.temperature === "number" ? parameters.temperature : undefined,
 			top_p: typeof parameters?.top_p === "number" ? parameters.top_p : undefined,

src/lib/server/textGeneration/types.ts

@@ -17,6 +17,8 @@ export interface TextGenerationContext {
 	forceMultimodal?: boolean;
 	/** Force-enable tool calling even if model does not advertise support */
 	forceTools?: boolean;
+	/** Inference provider preference: "auto", "fastest", "cheapest", or a specific provider name */
+	provider?: string;
 	locals: App.Locals | undefined;
 	abortController: AbortController;
 }

src/lib/stores/settings.ts

@@ -13,10 +13,11 @@ type SettingsStore = {
 	customPrompts: Record<string, string>;
 	multimodalOverrides: Record<string, boolean>;
 	toolsOverrides: Record<string, boolean>;
+	hidePromptExamples: Record<string, boolean>;
+	providerOverrides: Record<string, string>;
 	recentlySaved: boolean;
 	disableStream: boolean;
 	directPaste: boolean;
-	hidePromptExamples: Record<string, boolean>;
 	billingOrganization?: string;
 };
 

src/lib/types/Settings.ts

@@ -33,6 +33,13 @@ export interface Settings extends Timestamps {
 	 */
 	hidePromptExamples?: Record<string, boolean>;
 
+	/**
+	 * Per-model inference provider preference.
+	 * Values: "auto" (default), "fastest", "cheapest", or a specific provider name (e.g., "together", "sambanova").
+	 * The value is appended to the model ID when making inference requests (e.g., "model:fastest").
+	 */
+	providerOverrides?: Record<string, string>;
+
 	disableStream: boolean;
 	directPaste: boolean;
 
@@ -52,6 +59,7 @@ export const DEFAULT_SETTINGS = {
 	multimodalOverrides: {},
 	toolsOverrides: {},
 	hidePromptExamples: {},
+	providerOverrides: {},
 	disableStream: false,
 	directPaste: false,
 } satisfies SettingsEditable;

src/routes/+layout.ts

@@ -45,6 +45,10 @@ export const load = async ({ depends, fetch, url }) => {
 			welcomeModalSeenAt: settings.welcomeModalSeenAt
 				? new Date(settings.welcomeModalSeenAt)
 				: null,
+			// Ensure providerOverrides has a default value (may not exist in older DB records)
+			providerOverrides:
+				(settings as unknown as { providerOverrides?: Record<string, string> }).providerOverrides ??
+				{},
 		},
 		publicConfig: getConfigManager(publicConfig),
 		...featureFlags,

src/routes/conversation/[id]/+server.ts

@@ -1,5 +1,6 @@
 import { authCondition } from "$lib/server/auth";
 import { collections } from "$lib/server/database";
+import { config } from "$lib/server/config";
 import { models, validModelIdSchema } from "$lib/server/models";
 import { ERROR_MESSAGES } from "$lib/stores/errors";
 import type { Message } from "$lib/types/Message";
@@ -556,6 +557,11 @@ export async function POST({ request, locals, params, getClientAddress }) {
 					forceMultimodal: Boolean(userSettings?.multimodalOverrides?.[model.id]),
 					// Force-enable tools if user settings say so for this model
 					forceTools: Boolean(userSettings?.toolsOverrides?.[model.id]),
+					// Inference provider preference (HuggingChat only, skip for router models)
+					provider:
+						config.isHuggingChat && !model.isRouter
+							? userSettings?.providerOverrides?.[model.id]
+							: undefined,
 					locals,
 					abortController: ctrl,
 				};

src/routes/settings/(nav)/+layout.svelte

@@ -6,12 +6,18 @@
 	import { useSettingsStore } from "$lib/stores/settings";
 	import IconOmni from "$lib/components/icons/IconOmni.svelte";
 	import IconBurger from "$lib/components/icons/IconBurger.svelte";
+	import IconFast from "$lib/components/icons/IconFast.svelte";
+	import IconCheap from "$lib/components/icons/IconCheap.svelte";
 	import CarbonClose from "~icons/carbon/close";
 	import CarbonTextLongParagraph from "~icons/carbon/text-long-paragraph";
 	import CarbonChevronLeft from "~icons/carbon/chevron-left";
 	import LucideImage from "~icons/lucide/image";
 	import LucideHammer from "~icons/lucide/hammer";
 	import IconGear from "~icons/bi/gear-fill";
+	import { PROVIDERS_HUB_ORGS } from "@huggingface/inference";
+	import { usePublicConfig } from "$lib/utils/PublicConfig.svelte";
+
+	const publicConfig = usePublicConfig();
 
 	import type { LayoutData } from "../$types";
 	import { browser } from "$app/environment";
@@ -200,6 +206,42 @@
 						</span>
 					{/if}
 
+					{#if publicConfig.isHuggingChat && !model.isRouter && $settings.providerOverrides?.[model.id] && $settings.providerOverrides[model.id] !== "auto"}
+						{@const providerOverride = $settings.providerOverrides[model.id]}
+						{@const hubOrg =
+							PROVIDERS_HUB_ORGS[providerOverride as keyof typeof PROVIDERS_HUB_ORGS]}
+						{#if providerOverride === "fastest"}
+							<span
+								title="Provider: {providerOverride}"
+								class="grid size-[21px] flex-none place-items-center rounded-md bg-green-500/10 text-green-600 dark:text-green-500"
+								aria-label="Provider: {providerOverride}"
+								role="img"
+							>
+								<IconFast classNames="size-3" />
+							</span>
+						{:else if providerOverride === "cheapest"}
+							<span
+								title="Provider: {providerOverride}"
+								class="grid size-[21px] flex-none place-items-center rounded-md bg-blue-500/10 text-blue-600 dark:text-blue-500"
+								aria-label="Provider: {providerOverride}"
+								role="img"
+							>
+								<IconCheap classNames="size-3" />
+							</span>
+						{:else if hubOrg}
+							<span
+								title="Provider: {providerOverride}"
+								class="flex size-[21px] flex-none items-center justify-center rounded-md bg-gray-500/10 p-[0.225rem]"
+							>
+								<img
+									src="https://huggingface.co/api/avatars/{hubOrg}"
+									alt={providerOverride}
+									class="size-full rounded"
+								/>
+							</span>
+						{/if}
+					{/if}
+
 					{#if $settings.customPrompts?.[model.id]}
 						<CarbonTextLongParagraph
 							class="size-6 rounded-md border border-gray-300 p-1 text-gray-800 dark:border-gray-600 dark:text-gray-200"

src/routes/settings/(nav)/+server.ts

@@ -16,6 +16,7 @@ export async function POST({ request, locals }) {
 			customPrompts: z.record(z.string()).default({}),
 			multimodalOverrides: z.record(z.boolean()).default({}),
 			toolsOverrides: z.record(z.boolean()).default({}),
+			providerOverrides: z.record(z.string()).default({}),
 			disableStream: z.boolean().default(false),
 			directPaste: z.boolean().default(false),
 			hidePromptExamples: z.record(z.boolean()).default({}),