feat: add support for the Gemma-3 and Gemini Robotics models via the Gemini API

Author: techstrom

packages/types/src/providers/gemini.ts

@@ -234,4 +234,44 @@ export const geminiModels = {
 		supportsReasoningBudget: true,
 		maxThinkingTokens: 24_576,
 	},
+
+	// Robotics models
+	"gemini-robotics-er-1.5-preview": {
+		maxTokens: 8192,
+		contextWindow: 1_000_000,
+		supportsImages: true,
+		supportsNativeTools: true,
+		supportsPromptCache: false,
+		inputPrice: 0.3,
+		outputPrice: 2.5,
+	},
+
+	// Gemma 3 models
+	"gemma-3-27b-it": {
+		maxTokens: 8192,
+		contextWindow: 128_000,
+		supportsImages: true,
+		supportsNativeTools: false,
+		supportsPromptCache: false,
+		inputPrice: 0.07,
+		outputPrice: 0.5,
+	},
+	"gemma-3-12b-it": {
+		maxTokens: 8192,
+		contextWindow: 128_000,
+		supportsImages: true,
+		supportsNativeTools: false,
+		supportsPromptCache: false,
+		inputPrice: 0.03,
+		outputPrice: 0.1,
+	},
+	"gemma-3-4b-it": {
+		maxTokens: 8192,
+		contextWindow: 128_000,
+		supportsImages: true,
+		supportsNativeTools: false,
+		supportsPromptCache: false,
+		inputPrice: 0.02,
+		outputPrice: 0.07,
+	},
 } as const satisfies Record<string, ModelInfo>

src/api/providers/gemini.ts

@@ -162,8 +162,30 @@ export class GeminiHandler extends BaseProvider implements SingleCompletionHandl
 			? (this.options.modelTemperature ?? info.defaultTemperature ?? 1)
 			: info.defaultTemperature
 
+		// Check if the model is a Gemma 3 model
+		const isGemma3 = model.includes("gemma-3")
+
+		// Prepend system instruction to the first user message if it's a Gemma 3 model,
+		// as they don't support the system instruction parameter.
+		if (isGemma3 && systemInstruction) {
+			if (contents.length > 0 && contents[0].role === "user") {
+				const firstMessage = contents[0]
+				// Create a new text part for the system instruction
+				const systemPart = { text: systemInstruction }
+				// Prepend it to the existing parts
+				firstMessage.parts = [systemPart, ...firstMessage.parts]
+			} else {
+				// If no messages or first message is not user (e.g. starts with model),
+				// prepend a new user message with the system instruction.
+				contents.unshift({
+					role: "user",
+					parts: [{ text: systemInstruction }],
+				})
+			}
+		}
+
 		const config: GenerateContentConfig = {
-			systemInstruction,
+			...(isGemma3 ? {} : { systemInstruction }),
 			httpOptions: this.options.googleGeminiBaseUrl ? { baseUrl: this.options.googleGeminiBaseUrl } : undefined,
 			thinkingConfig,
 			maxOutputTokens,
@@ -351,6 +373,19 @@ export class GeminiHandler extends BaseProvider implements SingleCompletionHandl
 			TelemetryService.instance.captureException(apiError)
 
 			if (error instanceof Error) {
+				// Parse "limit: 15000, model: gemma-3-27b\nPlease retry in 31.714908887s"
+				// Use a more flexible regex to handle potential variations in spacing or text
+				const match = error.message.match(/limit:\s*(\d+)\s*,\s*model:\s*([^,\n]+).*?retry in\s*([\d.]+)/s)
+				if (match) {
+					const [, limit, model, retry] = match
+					throw new Error(
+						t("common:errors.gemini.resource_exhausted", {
+							limit,
+							model: model.trim(),
+							retry: Math.round(Number(retry)),
+						}),
+					)
+				}
 				throw new Error(t("common:errors.gemini.generate_stream", { error: error.message }))
 			}
 

src/i18n/locales/en/common.json

@@ -109,7 +109,8 @@
 			"thinking_complete_no_output": "(Thinking complete, but no output was generated.)",
 			"thinking_complete_truncated": "(Thinking complete, but output was truncated due to token limit.)",
 			"thinking_complete_safety": "(Thinking complete, but output was blocked due to safety settings.)",
-			"thinking_complete_recitation": "(Thinking complete, but output was blocked due to recitation check.)"
+			"thinking_complete_recitation": "(Thinking complete, but output was blocked due to recitation check.)",
+			"resource_exhausted": "Rate limit reached for {{model}}. The limit is {{limit}} requests per minute. Please retry in {{retry}} seconds."
 		},
 		"cerebras": {
 			"authenticationFailed": "Cerebras API authentication failed. Please check your API key is valid and not expired.",