Now LLM Service updates

  • Release version: Australia
  • Updated May 27, 2026
  • 5 minutes to read
    • Summarize
    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 Now LLM Service updates

    The Now LLM Service offers access to specialized large language models (LLMs) developed or enhanced by ServiceNow, as well as selected open-source models from the ServiceNow community and partners. These models support various AI-driven language tasks such as text generation, summarization, conversational AI, and content moderation within ServiceNow workflows. Model cards provide detailed information about each model’s context, intended use, training data, and limitations, helping customers choose the right model for their use cases.

    Show full answer Show less

    Key Features

    • Model Variety and Use Cases: Models include ServiceNow large and small language models (V1 and V2), third-party LLMs, Voice AI models for speech conversion, content moderation models, and specialized models like Inferred CSAT prediction.
    • Multi-language Support: The service supports multiple languages including English, German, French, Japanese, Dutch, French Canadian, Spanish, Brazilian Portuguese, and Italian to accommodate global teams.
    • JSON Output Format: Provides deterministic JSON responses to enable easier integration with applications, reduce errors, and optimize token consumption for high-frequency use cases.
    • Enhanced Instruction Following: Continual improvements in the models’ ability to accurately interpret and adhere to user instructions for more precise and actionable outputs.
    • Context Window Expansion: Increased context window from 16K to 32K tokens in recent model versions, enabling better understanding and coherence for long-form inputs.
    • Model Deployment Consolidation: Integration of multiple ServiceNow-related AI tasks into singular models reduces system complexity and improves overall performance.
    • Content Moderation and Security: Models like the Now Assist Guardian provide content moderation and guard against prompt injection and offensive content.

    Recent Updates and Practical Implications

    • June 2026 Update: Mandatory update to third-party default model versions (GAIC 13.1.2), change in GPT-5 Mini’s default reasoningeffort from “none” to “minimal” requiring testing and configuration checks, retirement of Claude Sonnet 4.0 (automatically redirected to Claude 4.5), and retirement of the Now LLM long-term support SKU with no required customer action.
    • May 2025 Release: Introduction of an advanced 12B parameter small language model fine-tuned for ServiceNow-specific tasks like Agent Assist and Text-to-Code, featuring improved instruction adherence, multilingual proficiency (notably Japanese), and extended support for workflow-related capabilities.
    • March 2025 Release: Deployment of a powerful 12B small language model optimized for case summarization, chat summarization, resolution notes, and knowledge base generation with enhanced text-to-code and content moderation capabilities.
    • November 2024 Enhancements: Added eight new supported languages, enabled JSON output for better integration, improved response determinism and error reduction, and enhanced instruction following for faster, more reliable results.

    What This Means for ServiceNow Customers

    ServiceNow customers can leverage the Now LLM Service to enhance automation, natural language understanding, and content generation within their workflows. The availability of multiple specialized models and continual improvements ensures that customers can select and deploy models tailored to their enterprise needs, with support for multilingual teams and streamlined integration via JSON. Regular updates and deprecations require customers to stay informed about version changes and configuration adjustments, such as the GPT-5 Mini reasoning effort setting, to maintain optimal performance and compatibility.

    The Now LLM Service provides access to specialized large language models (LLMs) that are developed by ServiceNow. It also provides access to open-source LLMs that are selected, configured, or enhanced by ServiceNow, from the ServiceNow community and partners. Review these reference materials and model cards for additional information about the Now LLM Service and about the models 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, such as 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 Now Assist Admin console and review the LLM service column.

    Model card for ServiceNow large language model
    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)

    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
    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)

    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
    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
    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 Now Assist Guardian model
    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
    This model is designed to ingest a conversation and predict a CSAT score as well as factors that explain the predicted score.

    June 2026

    The June release includes updates to third-party model defaults, 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

    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, maintain coherence over extended interactions, and support 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.

    March 2025

    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.

    November 2024

    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.
    • 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.