chore: import upstream snapshot with attribution

integrations/cerebras/eslint.config.mjs

@@ -0,0 +1,13 @@
+import rootConfig from '../../eslint.config.mjs'
+
+export default [
+  ...rootConfig,
+  {
+    languageOptions: {
+      parserOptions: {
+        project: ['./tsconfig.json'],
+        tsconfigRootDir: import.meta.dirname,
+      },
+    },
+  },
+]

integrations/cerebras/hub.md

@@ -0,0 +1,5 @@
+# Cerebras Integration
+
+This integration allows your bot to choose from a curated list of models from [Cerebras](https://cerebras.ai/developers/) for content generation and chat completions.
+
+Usage is charged to the AI Spend of your workspace in Botpress Cloud at the same pricing (at cost) as directly with Cerebras.

integrations/cerebras/integration.definition.ts

@@ -0,0 +1,31 @@
+import { IntegrationDefinition, z } from '@botpress/sdk'
+import { modelId } from 'src/schemas'
+import llm from './bp_modules/llm'
+
+export default new IntegrationDefinition({
+  name: 'cerebras',
+  title: 'Cerebras',
+  description:
+    'Get access to a curated list of Cerebras models for content generation and chat completions within your bot.',
+  version: '9.0.0',
+  readme: 'hub.md',
+  icon: 'icon.svg',
+  entities: {
+    modelRef: {
+      schema: z.object({
+        id: modelId,
+      }),
+    },
+  },
+  secrets: {
+    CEREBRAS_API_KEY: {
+      description: 'Cerebras API key',
+    },
+  },
+  attributes: {
+    category: 'AI Models',
+    repo: 'botpress',
+  },
+}).extend(llm, ({ entities: { modelRef } }) => ({
+  entities: { modelRef },
+}))

integrations/cerebras/package.json

@@ -0,0 +1,24 @@
+{
+  "name": "@botpresshub/cerebras",
+  "scripts": {
+    "build": "bp add -y && bp build",
+    "check:type": "tsc --noEmit",
+    "check:bplint": "bp lint",
+    "test": "vitest --run"
+  },
+  "private": true,
+  "dependencies": {
+    "@botpress/client": "workspace:*",
+    "@botpress/common": "workspace:*",
+    "@botpress/sdk": "workspace:*",
+    "openai": "^5.12.1"
+  },
+  "devDependencies": {
+    "@botpress/cli": "workspace:*",
+    "@botpress/sdk": "workspace:*",
+    "@botpresshub/llm": "workspace:*"
+  },
+  "bpDependencies": {
+    "llm": "../../interfaces/llm"
+  }
+}

integrations/cerebras/src/index.ts

@@ -0,0 +1,143 @@
+import { llm } from '@botpress/common'
+import { validateGptOssReasoningEffort } from '@botpress/common/src/llm/openai'
+import OpenAI from 'openai'
+import { DEFAULT_MODEL_ID, ModelId } from './schemas'
+import * as bp from '.botpress'
+
+const cerebrasClient = new OpenAI({
+  baseURL: 'https://api.cerebras.ai/v1',
+  apiKey: bp.secrets.CEREBRAS_API_KEY,
+})
+
+const languageModels: Record<ModelId, llm.ModelDetails> = {
+  // Reference:
+  // https://inference-docs.cerebras.ai/models/overview
+  // https://www.cerebras.ai/pricing
+  'gpt-oss-120b': {
+    name: 'GPT-OSS 120B (Preview)',
+    description:
+      'gpt-oss-120b is a high-performance, open-weight language model designed for production-grade, general-purpose use cases. It excels at complex reasoning and supports configurable reasoning effort, full chain-of-thought transparency for easier debugging and trust, and native agentic capabilities for function calling, tool use, and structured outputs.',
+    tags: ['preview', 'general-purpose', 'reasoning'],
+    input: {
+      costPer1MTokens: 0.25,
+      maxTokens: 131_000,
+    },
+    output: {
+      costPer1MTokens: 0.69,
+      maxTokens: 16_000,
+    },
+  },
+  'qwen-3-32b': {
+    name: 'Qwen3 32B',
+    description:
+      'Qwen3-32B is a world-class reasoning model with comparable quality to DeepSeek R1 while outperforming GPT-4.1 and Claude Sonnet 3.7. It excels in code-gen, tool-calling, and advanced reasoning, making it an exceptional model for a wide range of production use cases. NOTE: This model always uses thinking tokens (reasoning) by default, but we have configured it to avoid reasoning (not guaranteed) if the `reasoningEffort` parameter is not set. If the `reasoningEffort` parameter is set, the model will use thinking tokens. The model currently only supports "high" reasoning effort so any other value will be ignored.',
+    tags: ['general-purpose', 'reasoning'],
+    input: {
+      costPer1MTokens: 0.4,
+      maxTokens: 128_000,
+    },
+    output: {
+      costPer1MTokens: 0.8,
+      maxTokens: 16_000,
+    },
+  },
+  'llama-4-scout-17b-16e-instruct': {
+    name: 'Llama 4 Scout 17B',
+    description:
+      'Llama 4 Scout 17B Instruct (16E) is a mixture-of-experts (MoE) language model developed by Meta, uses 16 experts per forward pass, activating 17 billion parameters out of a total of 109B. It supports native multimodal input (text and image) and multilingual output (text and code) across 12 supported languages. Designed for assistant-style interaction and visual reasoning, it is instruction-tuned for use in multilingual chat, captioning, and image understanding tasks.',
+    tags: ['general-purpose'],
+    input: {
+      costPer1MTokens: 0.65,
+      maxTokens: 32_000,
+    },
+    output: {
+      costPer1MTokens: 0.85,
+      maxTokens: 16_000,
+    },
+  },
+  'llama3.1-8b': {
+    name: 'Llama 3.1 8B',
+    description:
+      'Meta developed and released the Meta Llama 3 family of large language models (LLMs), a collection of pretrained and instruction tuned generative text models in 8B and 70B sizes. The Llama 3 instruction tuned models are optimized for dialogue use cases and outperform many of the available open source chat models on common industry benchmarks.',
+    tags: ['low-cost', 'general-purpose'],
+    input: {
+      costPer1MTokens: 0.1,
+      maxTokens: 32_000,
+    },
+    output: {
+      costPer1MTokens: 0.1,
+      maxTokens: 16_000,
+    },
+  },
+  'llama3.3-70b': {
+    name: 'Llama 3.3 70B',
+    tags: ['general-purpose'],
+    description:
+      'Meta developed and released the Meta Llama 3 family of large language models (LLMs), a collection of pretrained and instruction tuned generative text models in 8B and 70B sizes. The Llama 3 instruction tuned models are optimized for dialogue use cases and outperform many of the available open source chat models on common industry benchmarks.',
+    input: {
+      costPer1MTokens: 0.85,
+      maxTokens: 128_000,
+    },
+    output: {
+      costPer1MTokens: 1.2,
+      maxTokens: 16_000,
+    },
+  },
+}
+
+const provider = 'Cerebras'
+
+export default new bp.Integration({
+  register: async () => {},
+  unregister: async () => {},
+  actions: {
+    generateContent: async ({ input, logger, metadata }) => {
+      const output = await llm.openai.generateContent<ModelId>(
+        <llm.GenerateContentInput>input,
+        cerebrasClient as any, // TODO: fix mismatch of openai version
+        logger,
+        {
+          provider,
+          models: languageModels,
+          defaultModel: DEFAULT_MODEL_ID,
+          overrideRequest: (request) => {
+            if (input.model?.id === 'qwen-3-32b' && input.reasoningEffort === undefined) {
+              // As per the Cerebras documentation, the Qwen-3-32B model uses thinking tokens by default but we can suggest it to not do reasoning by appending `/no_think` to the prompt.
+              for (const message of request.messages) {
+                if (
+                  message.role === 'user' &&
+                  typeof message.content === 'string' &&
+                  !message.content.endsWith('/no_think')
+                ) {
+                  message.content += ' /no_think'
+                  break
+                }
+              }
+            }
+
+            if (input.model?.id === 'gpt-oss-120b') {
+              request.reasoning_effort = validateGptOssReasoningEffort(input, logger)
+
+              // GPT-OSS models don't work well with a stop sequence, so we have to remove it from the request.
+              delete request.stop
+
+              // Reasoning models don't support temperature
+              delete request.temperature
+            }
+
+            return request
+          },
+        }
+      )
+      metadata.setCost(output.botpress.cost)
+      return output
+    },
+    listLanguageModels: async ({}) => {
+      return {
+        models: Object.entries(languageModels).map(([id, model]) => ({ id: <ModelId>id, ...model })),
+      }
+    },
+  },
+  channels: {},
+  handler: async () => {},
+})

integrations/cerebras/src/schemas.ts

@@ -0,0 +1,10 @@
+import { z } from '@botpress/sdk'
+
+export const DEFAULT_MODEL_ID = 'llama3.1-8b'
+
+export const modelId = z
+  .enum(['gpt-oss-120b', 'qwen-3-32b', 'llama-4-scout-17b-16e-instruct', 'llama3.1-8b', 'llama3.3-70b'])
+  .describe('Model to use for content generation')
+  .placeholder(DEFAULT_MODEL_ID)
+
+export type ModelId = z.infer<typeof modelId>

integrations/cerebras/tsconfig.json

@@ -0,0 +1,8 @@
+{
+  "extends": "../../tsconfig.json",
+  "compilerOptions": {
+    "paths": { "*": ["./*"] },
+    "outDir": "dist"
+  },
+  "include": [".botpress/**/*", "definitions/**/*", "src/**/*", "*.ts"]
+}

integrations/cerebras/vitest.config.ts

@@ -0,0 +1,2 @@
+import config from '../../vitest.config'
+export default config