--- sourceDocument: Yokohama Enable AI sourceDocumentLink: https://www.servicenow.com/docs/r/yokohama/intelligent-experiences Release : - yokohama ft:locale : - en-US ft:publication_title : - Yokohama Enable AI ft:clusterId : - platai bundleId : - platai workflow : - Platform --- # Model provider updates # Model provider updates {#ariaid-title1} Release version: Yokohama Updated September 2, 2026 ![](https://www.servicenow.com/docs/portal-asset/ico-clock) 8 minutes to read Summarize ![AI sparkle icon](https://servicenow.com/docs/portal-asset/ai-sparkle-icon) Summarized using AI This content was generated using new OpenAI-powered functionality. Results are provided on an as is basis and are not guaranteed to be accurate or complete. ## Summary of Model provider updates This document outlines the updates to model providers and large language models (LLMs) available for ServiceNow skills and agents, focusing on releases from November 2024 through October 2026. It details model cards describing the context, intended use, training data, and limitations for each model. Updates include new model introductions, retirements, deprecations, and enhancements to improve AI-driven automation, natural language understanding, and conversational AI capabilities within ServiceNow workflows. Show full answer Show less ## Key Features * **Model Cards:** Provide detailed information about ServiceNow's proprietary and third-party LLMs, including ServiceNow large and small language models (both V1 and V2), third-party models, Voice AI models, AI Guardian for content moderation, and specialized models for CSAT prediction. * **Integrated Third-Party Model Providers:** Google Gemini, AWS Claude, and Azure OpenAI models are integrated and selectable in the AI Admin Hub, improving flexibility in choosing models tailored to instance configuration and geographic region. * **Model Provider Strategy Shift:** Default base system skills and agents are transitioning from the Now LLM Service to integrated third-party providers while preserving existing explicit model selections. * **New and Retired Models:** * October 2026 introduced the ServiceNow hosted Gemma 4 26B-A4B model and retired Gemini 2.5 models. * ServiceNow large and small language model V2 versions are deprecated for new development, with retirement planned in December 2026, replaced by Gemma. * September 2026 added restricted AWS Anthropic models, which cannot be selected for skills. * June 2026 updated third-party model defaults and changed the default reasoningeffort setting for GPT-5 Mini from none to minimal for improved performance. * May 2025 and March 2025 releases introduced advanced 12B general-purpose small language models optimized for ServiceNow workflows, including improved instruction adherence, larger context windows, multilingual support, and enhanced text-to-code and content moderation capabilities. * **JSON Output Support and Multilingual Enhancements:** The Now LLM Service supports structured JSON output for easier integration and reduced errors, along with expanded multilingual capabilities covering nine languages to support global teams. ## What Customers Can Expect * Enhanced model options allowing selection of best-fit LLMs for automation, case summarization, conversational AI, content generation, and safety moderation within their ServiceNow environment. * Ongoing improvements in model accuracy, instruction adherence, and efficiency, enabling more reliable and contextually aware AI-driven workflows. * Flexibility to continue using existing models without disruption while planning for upcoming retirements and migrations to newer models like Gemma. * Clear guidance on model provider configurations through the AI Admin Hub and AI Control Tower, enabling easy management of model selections per instance and region. * Awareness of deprecated models and settings requiring regression testing and updates to align with new defaults, such as the reasoningeffort parameter for GPT-5 Mini. By understanding these updates, ServiceNow customers can effectively manage AI model selections, optimize AI capabilities in workflows, and maintain alignment with the evolving ServiceNow AI ecosystem. Review these updates to learn which model providers and models are available for your skills and agents. Review the model cards for information about how each model is intended to be used. ## Model cards Large language models (LLMs) are complex machine-learning models that are trained on large datasets like websites and documentation to perform language-related tasks. Examples include text generation for case summaries and resolution notes. Model cards explain the specific model's context, intended use, training data, limitations, and other important information. These model cards are for skills that use the Now LLM Service. There are certain skills, such as Now Assist Multi-Turn Catalog Ordering, that use Azure OpenAI instead. To see what LLM a skill is using, you can check the skill list in the AI Admin Hub console and review the LLM service column. [Model card for ServiceNow large language model](https://downloads.docs.servicenow.com/resource/enus/infocard/sn-llm.pdf) : Model used for AI-driven solutions to support natural language understanding, automation, and decision support. : This model card is available in Yokohama patch 1 and later. [Model card for ServiceNow large language model (V2)](https://downloads.docs.servicenow.com/resource/enus/infocard/sn-llm-v2.pdf) : Model for enterprise AI that enhances text-based automation and content generation in ServiceNow workflows, including requester OOTB skills, custom skills, and agentic use cases. This model is for Generative AI Controller application 11.2 or higher. [Model card for ServiceNow small language model](https://downloads.docs.servicenow.com/resource/enus/infocard/sn-slm.pdf) : Model used for enterprise AI applications by enhancing text-based automation and content generation within ServiceNow workflows. : This model card is available in Yokohama patch 1 and later. [Model card for ServiceNow small language model (V2)](https://downloads.docs.servicenow.com/resource/enus/infocard/sn-slm-v2.pdf) : Model for enterprise AI that enhances text-based automation and content generation in ServiceNow workflows, including creator and fulfiller OOTB skills as well as custom skills. This model is for Generative AI Controller application 11.2 or higher. [Model card for ServiceNow third party large language model](https://downloads.docs.servicenow.com/resource/enus/infocard/third-party-llm.pdf) : Model used for AI-driven solutions for text generation, summarization, and conversational AI. [Model card for ServiceNow Voice AI Speech-to-Text and Text-to-Speech models](https://downloads.docs.servicenow.com/resource/enus/infocard/sn-voice.pdf) : Models used within ServiceNow AI Voice Agents for converting spoken user input to text and generating natural-sounding speech from AI responses. [Model card for ServiceNow AI Guardian model](https://downloads.docs.servicenow.com/resource/enus/infocard/sn-na-guardian.pdf) : This model provides content moderation and helps identify different kinds of prompt injection attacks and offensive content. [Model card for ServiceNow Inferred CSAT and Factors large language model](https://downloads.docs.servicenow.com/resource/enus/infocard/csat-llm.pdf) : This model is designed to ingest a conversation and predict a CSAT score as well as factors that explain the predicted score. ## October 2026 {#now-llm-model-updates__section_october_2026} The October release adds a ServiceNow hosted model. It also retires the Gemini 2.5 models and begins the deprecation of the Now LLM Service V2 models. * New ServiceNow hosted model: Gemma 4 26B-A4B is available if you require a hosted or self-hosted model. Third-party models remain the default. You can select Gemma for your skills and agents instead of the default. * Gemini 2.5 models retired: Gemini 2.5 Pro, Gemini 2.5 Flash, and Gemini 2.5 Flash Lite are retired in the October release. Review your model selections and choose a supported Gemini model. * V2 models deprecated for new development: The ServiceNow large language model (V2) and small language model (V2) are deprecated for new development. Your existing skills and agents continue to work. Retirement is planned for the December release, when these models are replaced by Gemma. {#now-llm-model-updates__ul_october_2026} ## September 2026 {#now-llm-model-updates__section_september_2026} The September release adds restricted models from AWS Anthropic. Restricted models support specific features only. You can't select them for your skills and agents. This release doesn't add any general purpose models. The models that you can select in the AI Admin Hub console are unchanged. * Claude Sonnet 5 and Claude Fable 5 support the Build Agent. * Restricted replaces limited: The card now uses Restricted instead of Limited to identify models that you can't configure. Claude Opus 4.6 and Gemini 2.5 Flash Lite use the new term. Their supported features haven't changed. {#now-llm-model-updates__ul_september_2026} ## July 2026 {#now-llm-model-updates__section_july_2026} The Now LLM Service model strategy has moved toward integrated model provider flexibility. Store application updates have started shipping, and the default model for base system skills and agents is changing from the Now LLM Service to an integrated third-party model provider. If you haven't changed the default model for a skill, that skill uptakes the new default. * Integrated third-party model providers: The integrated third-party model providers are Google Gemini, AWS Claude, and Azure OpenAI. * New general purpose models: You can select these models for your skills and agents in the AI Admin Hub console. Model availability varies by your instance, your configuration, and your geographic region. * Azure OpenAI: GPT-5.1, GPT-5.4 Mini, and GPT-5.1 Mini * Google Gemini: Gemini 3.5 Flash * The Now LLM Service is still supported: The change to the default model provider doesn't remove access to the Now LLM Service. You can still select it for your skills and agents if you require a ServiceNow hosted model. * Existing selections preserved: If you have already selected a model provider, that configuration is preserved. The update doesn't overwrite an explicit selection. * Changing your model provider: To select a model provider other than the default, review the providers that are allowed for your instance in AI Control Tower. Navigate to ConfigurationsControlsAI model providers, and then make your changes in the ServiceNow Otto Admin Center by navigating to SettingsManage AI ModelsManage Model Providers. * Explicit model names required: Generic model references such as `cloud_small` and `cloud_large` are no longer supported as of December 2026, when the Now LLM Service transitions to integrated third-party model providers. {#now-llm-model-updates__ul_july_2026} ## June 2026 {#now-llm-model-updates__section_june_2026} The June release includes updates to third-party model defaults. It also contains a change to the default reasoning effort setting for GPT-5 Mini, and the retirement of the Now LLM long-term support (LTS) SKU. * Third-party default model version update: Teams that did not update their third-party default model versions to the latest available versions in the May release must do so in the June release. GAIC 13.1.2 is the required version for this update. * GPT-5 Mini: `reasoning_effort` default change: The default `reasoning_effort` setting for GPT-5 Mini has changed from `none` to `minimal`. This change is included in GAIC Snapshot 14.0.0, which is compatible with Now Assist for Platform 12.0.0. Teams using GPT-5 Mini should run regression and functional testing to confirm that the new default works as expected. If you explicitly set `reasoning_effort` in your generative AI config additional properties, smoke test to verify there are no unexpected effects. If you have `reasoning_effort: none` set in additional properties, update the value to `minimal` and run regression and functional testing. * Claude Sonnet 4.0 retirement: Claude Sonnet 4.0 references are being redirected on the backend. No team action is required. * `claude_large` / Claude 4.0 Sonnet redirects to Claude 4.5 Sonnet * `claude_small` / Claude 4.0 Sonnet redirects to Claude 4.5 Sonnet * Now LLM LTS SKU retired: ServiceNow no longer offers the LTS model SKU. There are no testing requirements or expectations for teams related to this retirement. ## May 2025 {#now-llm-model-updates__section_qmv_nns_jfc} An advanced 12B general-purpose small language model (SLM) with a singular, high-performance architecture that supports a wide range of tasks in ServiceNow's context was released. Fine-tuned on Mistral-Nemo-12B-Instruct, this model is designed and optimized for tasks like Agent Assist, Text-to-Flow, Text-to-Cypher, Safety \& Content Moderation and Text-to-Code. Key Enhancements: * Enhanced instruction adherence: Improved the model's capability to accurately interpret and follow user instructions, ensuring that the model can better understand and execute complex commands. Leading to more precise and reliable outcomes than previous releases. * Increased context window: Increased context window from 16K to 32K, enabling the model to better understand long-form inputs. This increase maintains coherence over extended interactions, and supports more complex tasks with richer contextual awareness. * Improved multilingual proficiency: Boosted performance across languages compared to previous releases, with notable enhancements in Japanese processing. * Optimized for ServiceNow workflow related capabilities: Extended support coverage for Text-to-Flow, and improved the performance of Text-to-Code, Text-to-Cypher etc. * Continuously enhanced model deployment consolidation: Integrates ServiceNow-related tasks into a single model, reducing system complexity at the same time while elevating overall performance. {#now-llm-model-updates__ul_y25_sns_jfc} ## March 2025 {#now-llm-model-updates__section_yks_3cn_m2c} A powerful 12B general-purpose small language model (SLM) designed to enhance a wide range of applications, including text-to-code and agent use cases was released. Fine-tuned on Mistral-Nemo-12B, it streamlines deployment and consolidates multiple functionalities into a singular, architecture. Key Enhancements: * Optimized to fulfill use cases: Enhances case summarization, chat summarization, resolution notes, and knowledge base generation across supported languages, including improvements in Japanese quality. * Superior text-to-code and text-to-cypher performance: Delivers major advancements in Glide JavaScript and generic JavaScript editing and generation, along with improved accuracy in query generation and execution for structured databases. * Robust content moderation and safety: Provides stronger protection against adversarial prompts, jail-breaking attempts, and harmful content generation, ensuring safer deployment with built-in content filtering. * Unified model deployment:integrates ServiceNow-related tasks into a single model, thereby reducing system complexity while elevating overall performance. * Improved instruction adherence: Delivers better instruction following and consistency across varying levels of prompt and instruction strictness than the current text-to-text NowLLM. {#now-llm-model-updates__ul_xv4_1q5_m2c} ## November 2024 {#now-llm-model-updates__section_o24_ym3_gdc} Several key improvements were added to the Now LLM Service that are aimed at enhancing performance and quality. * Multilingual support: Now LLM Service supports 8 additional languages, enabling global teams to use the model in their native languages. The supported languages are: English, German, French, Japanese, Dutch, French Canadian, Spanish, Brazilian Portuguese, and Italian. * JSON format support: The model now provides output in JSON format, making it easier for developers to integrate with various applications and automate workflows seamlessly. * Deterministic responses: JSON mode ensures structured, consistent output, which improves predictability and reliability when integrating with applications. * Error reduction: Unlike free-form text mode, JSON responses are less prone to format errors or stray characters, minimizing integration issues. * Lower token consumption: The fixed structure of JSON can reduce token usage, making it more efficient and cost-effective for applications with high response frequency. {#now-llm-model-updates__ul_kmm_b43_gdc} * Improvements in instruction following: The model has been fine-tuned to understand and follow instructions more precisely. This enables the model to deliver more to-the-point and actionable responses, helping users get the information they need faster and more efficiently. {#now-llm-model-updates__ul_mwj_nn3_gdc} *[\>]: and then