Model

CLEVI models are divided into three product families: CIP, which handles general intelligence and work automation; Cosa, which handles speech generation and recognition; and Cova, which handles information extraction from documents and images. Each product can be used independently or connected into services covering document recognition, analysis, task execution and voice guidance.

CIP

CIP handles general intelligence and work automation. It consists of cip-5.5-im and cip-5.5-mm, called through a gateway, and cip-5.5-sm, deployed on edge servers and on-premises.

cip-5.5-im and cip-5.5-mm

cip-5.5-im is suited to tasks with clear objectives and scope. Typical tasks include extracting the required information from provided materials, producing outputs in a specified format, and modifying code to meet requirements. The maximum 64K output is useful for producing detailed reports or outputs consisting of multiple sections.

• Development assistance: Write functions and features, fix well-defined bugs, and implement changes in line with existing patterns.
• Test generation: Draft tests and validation cases based on requirements and implementations.
• Document processing: Handle summarisation, classification, item extraction, format conversion and drafting.
• Multimodal analysis: Review tables and charts in screen images and documents together with relevant descriptions.
• Interactive work support: Answer questions based on work materials and produce the required outputs.
• Bulk processing: Use for individual summaries of multiple documents, request classification and data conversion.
• Subagents: Handle clearly bounded subtasks such as code exploration, reviewing specific files and organising materials.

For example, it can be used to receive feature requirements and related code, then produce proposed changes, tests and a change description, or to classify submitted documents and create summaries for the relevant staff.

cip-5.5-mm is suited to tasks that require identifying relationships between information and maintaining consistency across the overall task. In development, it reviews the connections between requirements, design, implementation and testing; in document analysis, it synthesises the claims and evidence, differences and contradictions across multiple materials.

• Complex feature implementation: Change code, tests and documentation together while considering the contracts and behaviour of multiple modules.
• Large-scale code analysis: Review call relationships, data flows and the scope of change impacts.
• Incident root-cause analysis: Compare logs, configurations, code and execution results to organise hypotheses about the cause and procedures for verification.
• Design and decision support: Structure requirements and constraints, and analyse the impact of implementation alternatives.
• Multi-document synthesis: Organise commonalities, differences and contradictions in reports and technical documents based on evidence.
• Multimodal integrated review: Interpret text and visual materials together to analyse issues and requirements.
• Multi-step agents: Use for work involving research, planning, tool execution and result review.
• Professional document writing: Produce technical reports, design proposals and detailed review documents that reflect complex conditions and supporting evidence.

For example, it is used to review relevant codes and logs from errors occurring across multiple services together, or for tasks spanning requirements analysis, implementation and review of validation results for complex features.

cip-5.5-sm

cip-5.5-sm performs local inference and task automation on edge servers and in on-premises environments. As the model size can be adjusted within the 24B–40B range, its configuration can be selected to match the computing resources of the deployment equipment and the required performance on site. Because it processes data close to where it is generated, it is used to integrate with on-site systems and utilise internal data. By configuring the required models and tools locally, it can also operate in environments with limited external connectivity.

It interprets information generated on site and connects it to processing procedures. It classifies logs and requests, extracts the required information from data, and performs repetitive tasks using connected internal tools and scripts.

• On-site event processing: Classifies equipment logs and status information and selects events that require verification.
• Data preprocessing: Extracts key items from work records and classifies, tags and normalises them.
• Request routing: Passes received requests to the responsible system or processing procedure.
• Local task agent: Performs repetitive tasks by querying internal systems and executing scripts.
• On-premises task support: Answers questions based on internal materials and organises task details.
• Follow-up analysis support: Selects materials requiring further review and summarises the key points.

For example, it is used in a workflow that classifies operational logs on a site server, retrieves related history, and then prepares an incident summary for the person responsible.

Cosa

Cosa is a range of speech products that converts text to speech, recognises input speech as text, and registers and reuses desired voices. It is used for service guidance, content creation, accessibility support, and voice interfaces for conversational agents.

cosa-a and cosa-b handle speech generation, while cosa-asr handles speech recognition. cosa-a and cosa-b are separate models that provide their own voice lists and synthesis settings, and are selected through the same service interface. The features commonly provided by both models are as follows.

• Text-to-speech: Synthesises the entered text in the selected voice.
• Voice cloning: Registers a voice based on a reference recording for which you have permission to use it, then synthesises new sentences.
• Voice design: Creates and registers a voice by describing the characteristics of the desired voice.
• Voice reuse: Uses a registered voice for subsequent synthesis.
• Synthesis settings: Adjusts the speaking speed, language and other settings.
• Streaming delivery: Synthesises text phrase by phrase as it arrives and delivers the audio sequentially.
• Playback control: Supports pausing, resuming, stopping and entering additional sentences.

With incremental text input, voice playback can start before the entire response is complete. While CIP is composing the response, you can configure a service in which Cosa reads out the phrases that are ready.

cosa-a

cosa-a supports text-to-speech, voice cloning, voice design and streaming output. In addition to basic functions for selecting the voice, language and speaking speed, it provides additional controls, including settings related to the number of synthesis steps, guidance strength and generation length. It is used to adjust settings and review results during voice production, or to apply multiple settings to the same script and select the result.

• Synthesises announcements and notifications for apps, kiosks and business systems.
• Creates audio versions of educational materials and user guides.
• Produces narration for video and audio content.
• Creates recurring content using a registered voice.
• Used for streaming voice output from conversational agents.
• Used for voice production with detailed synthesis settings.

cosa-b

cosa-b supports preset voice selection, description-based voice design, reference-voice-based cloning and streaming output. You can choose a voice from your own voice library or register a new voice for use in a service.

Voice cloning uses a reference recording and the corresponding text. It also supports a process in which the reference text is generated through speech recognition during registration, and registered voices can be reused for subsequent synthesis.

• Configures the voice personas of services and characters.
• Creates custom voices from a description or reference recording.
• Synthesises brand announcements, character dialogue and content narration.
• Used for voice responses from conversational agents.
• Synthesises changing announcements and scripts in the same voice.

When choosing between cosa-a and cosa-b, use the desired voice, the actual synthesis results, and the controls you need as your criteria.

cosa-asr

cosa-asr converts input speech into text. It documents recorded material and converts voice requests into input that downstream business systems can process.

• Transcribe voice memos and recorded material.
• Convert voice-based questions and business requests into input.
• Generate reference text for voice cloning.
• Use it for processing that continues with summarisation, classification, and search after transcription.

Passing the transcription results to CIP leads to content summarisation, request classification, and business draft creation. Synthesising the results with cosa-a or cosa-b creates a business service with voice input and output.

Cova

cova-1 is an LLM-based model dedicated to vision OCR. It recognises text in documents and images, extracts layouts and table structures, and provides brief descriptions of visual elements.

It converts documents as viewed by people into a format that agents and business systems can use more easily. The body text and tables are extracted as structured content, while images included in the document are conveyed as brief descriptions, allowing downstream agents to understand both the document's content and its visual context. By structuring a workflow that first reviews the material using the extracted text and descriptions, then examines only the original sources requiring detailed confirmation, you can reduce the need to repeatedly input the original images and save context tokens.

• Document digitisation: Converts paper documents and scanned materials into text-based resources.
• Business document intake: Extracts content from application forms, supporting documents and reports.
• Table data processing: Uses tables included in documents for subsequent data processing.
• Pre-processing for knowledge search: Converts document images and visual materials into searchable text and structures.
• Document review automation: Delivers extraction results to CIP to verify items, create summaries and compare materials.
• Agent input optimisation: Organises the body text, tables and image descriptions of long documents to deliver only the necessary information.

For example, while extracting the body text and tables from a report, it describes an embedded graph as a graph showing quarterly sales trends. The agent uses this description to understand the graph's presence and purpose, and reviews the original further when it needs exact figures or trends.

Product connections

Connecting products enables you to configure the following flows.

• Document-based operations: Cova extracts the body text, tables and image descriptions, CIP selectively reviews the necessary originals to perform analysis and process tasks, and Cosa delivers the results by voice.
• Voice-based operations: Cosa ASR converts voice requests into text, and cosa-a or cosa-b reads out the answers processed by CIP.
• Field automation: cip-5.5-sm classifies field data and performs local tasks, while passing materials requiring additional comprehensive analysis to cip-5.5-im or cip-5.5-mm.

Calls allowed by the gateway

The Gateway's plans determine the calls permitted for each model. The table below shows the scope enabled by the LLM_FREE, TTS_FREE and STT_FREE plans, which are applied automatically without an application. Models enabled through other routes are not included.

cip-5.5-sm and cova-1 are not included in this plan table. cip-5.5-im-chat, cip-5.5-mm-h, cosa-tts, ivy-4-embedding, ivy-4-embedding-mm, and ivy-4-mm in the table are model IDs that are not included in the product descriptions above. The API Calling guide explains which path each call key corresponds to.