Models

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 on its own or connected into a service that runs from document recognition through analysis, task execution, and spoken guidance.

CIP

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

cip-5.5-im and cip-5.5-mm

cip-5.5-im suits tasks with a clear goal and scope. Typical work includes extracting the needed information from provided material, producing output in a specified format, and modifying code to match requirements. The 64K maximum output is used for detailed reports and multi-part deliverables.

• Development assistance: writes functions and features, fixes clearly scoped bugs, and implements to existing patterns.
• Test generation: writes test drafts and verification cases based on requirements and implementation.
• Document processing: handles summarization, classification, item extraction, format conversion, and drafting.
• Multimodal analysis: reviews screen images and the tables and charts in documents alongside related explanations.
• Conversational work support: answers questions based on work material and produces the needed output.
• Bulk processing: used for summarizing many documents individually, classifying requests, and converting data.
• Subagents: takes on clearly bounded subtasks such as code exploration, reviewing specific files, and organizing material.

For example, it can take a feature requirement and the related code and produce a proposed change, tests, and a change description, or classify incoming documents and produce a summary for the person in charge.

cip-5.5-mm suits tasks that require understanding the relationships between pieces of information and keeping the whole task consistent. In development, it reviews how requirements, design, implementation, and tests connect; in document analysis, it brings together the claims and evidence, differences, and contradictions across multiple sources.

• Complex feature implementation: changes code, tests, and documentation together, taking the contracts and behavior of multiple modules into account.
• Large-scale code analysis: reviews call relationships, data flow, and the scope of a change's impact.
• Incident root-cause analysis: cross-checks logs, configuration, code, and execution results to lay out cause hypotheses and verification steps.
• Design and decision support: structures requirements and constraints and analyzes the impact of implementation alternatives.
• Multi-document synthesis: organizes the commonalities, differences, and contradictions across reports and technical documents, grounded in evidence.
• Integrated multimodal review: interprets text and visual material together to analyze problems and requirements.
• Multi-step agents: used for work that runs from research and planning through tool execution and result review.
• Specialist document writing: writes technical reports, design proposals, and detailed review documents that reflect complex conditions and evidence.

For example, it can review the related code and logs for errors that occurred across multiple services together, or handle a complex feature end to end, from requirements analysis through implementation and review of verification results.

cip-5.5-sm

cip-5.5-sm runs local inference and work automation in edge-server and on-premises environments. Because the model size can be adjusted within the 24B to 40B range, you choose a configuration that fits the compute resources of the deployment hardware and the performance the site requires. It processes data close to where it is generated, which suits integration with on-site systems and use of internal data, and with the required models and tools set up locally, it can operate even in environments with restricted external connectivity.

Its role is to interpret information generated on site and route it into processing procedures. It classifies logs and requests, extracts the needed information from data, and performs repetitive work through connected internal tools and scripts.

• On-site event handling: classifies equipment logs and status information and flags events that need attention.
• Data preprocessing: extracts key items from work records and classifies, tags, and normalizes them.
• Request routing: forwards received requests to the responsible system or processing procedure.
• Local work agent: performs repetitive tasks by querying internal systems and running scripts.
• On-premises work support: answers questions based on internal material and organizes work content.
• Follow-up analysis support: selects material that needs further review and summarizes the key points.

For example, it is used in a flow where a site server classifies operational logs, looks up related history, and writes an incident summary for the person in charge.

Cosa

Cosa is the speech product family that converts text to speech, recognizes speech input as text, and registers voices for reuse. It is used for service announcements, content production, accessibility support, and the voice interface of conversational agents.

Speech generation is handled by cosa-a and cosa-b, and speech recognition by cosa-asr. cosa-a and cosa-b are separate models, each with its own voice list and synthesis settings, and you select between them in the same service interface. The two models share the following features.

• Text-to-speech: synthesizes the input sentences in the selected voice.
• Voice cloning: registers a voice from a reference recording you are permitted to use and synthesizes new sentences with it.
• Voice design: creates and registers a voice from a description of the desired voice characteristics.
• Voice reuse: uses a registered voice in later synthesis.
• Synthesis settings: adjusts speaking rate, language, and other settings.
• Streaming delivery: synthesizes phrase by phrase as text arrives and delivers the audio in sequence.
• Playback control: supports pause, resume, and stop, plus adding more sentences.

With incremental text input, speech playback can start before the full answer is complete. You can build a service in which Cosa reads out the phrases that are ready while CIP is still writing the answer.

cosa-a

cosa-a supports text-to-speech, voice cloning, voice design, and streaming output. In addition to the basic options of selecting a voice, language, and speaking rate, it provides extra controls such as the number of synthesis steps, guidance strength, and settings related to generation length. It is used in flows where you adjust settings and review the results during voice production, or apply several settings to the same script and pick the result.

• Synthesizes announcements and notifications for apps, kiosks, and business systems.
• Turns training materials and user manuals into audio.
• Produces narration for video and audio content.
• Produces recurring content with a registered voice.
• Used for streaming voice output in conversational agents.
• Used for voice production that fine-tunes synthesis settings.

cosa-b

cosa-b supports preset voice selection, description-based voice design, reference-based voice cloning, and streaming output. You pick a voice from its own voice list or register a new voice for your service.

Voice cloning uses a reference recording and its matching text. During registration, the reference text can also be generated by speech recognition, and a registered voice is reused in later synthesis.

• Builds voice personas for services and characters.
• Produces custom voices from a description or a reference recording.
• Synthesizes brand announcements, character lines, and content narration.
• Used for voice responses in conversational agents.
• Synthesizes announcement messages and scripts that change over time in the same voice.

When choosing between cosa-a and cosa-b, base the decision on the voice you want, the actual synthesis results, and the controls you need.

cosa-asr

cosa-asr converts speech input to text. It turns recordings into documents, or converts spoken requests into input that downstream business systems can process.

• Transcribes voice memos and recordings.
• Turns spoken questions and work requests into input.
• Generates reference text for voice cloning.
• Used for processing that continues from transcription into summarization, classification, and search.

Passing the transcript to CIP leads to content summarization, request classification, and drafting work. Synthesizing the result with cosa-a or cosa-b gives you a business service with both speech input and speech output.

Cova

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

It converts documents meant for people into a form that agents and business systems can use. Body text and tables are extracted as structured content and images in the document are passed on as short descriptions, so a downstream agent grasps both the document's content and its visual context. If you build a flow that first reads the extracted text and descriptions and then reviews only the originals that need a closer look, you reduce how often original images are fed in again and save context tokens.

• Document digitization: converts paper documents and scans into text-based material.
• Business document intake: extracts content from application forms, supporting documents, and reports.
• Table data processing: uses the tables in documents for downstream data processing.
• Knowledge search preprocessing: converts document images and visual material into searchable text and structure.
• Document review automation: passes extraction results to CIP to check items, summarize, and compare across materials.
• Agent input optimization: organizes the body text, tables, and image descriptions of long documents so that only the needed information is passed on.

For example, while extracting the body text and tables of a report, it describes an inserted graph as a graph showing quarterly revenue trends. From that description the agent learns that the graph exists and what it is for, and reviews the original when it needs exact figures or trends.

Connecting the product families

By connecting the products, you can build the following flows.

• Document-based work: Cova extracts body text, tables, and image descriptions; CIP selectively checks the originals it needs and performs analysis and task processing; Cosa delivers the results as speech.
• Speech-based work: Cosa ASR converts a spoken request to text, and cosa-a or cosa-b reads out the answer that CIP produced.
• On-site automation: cip-5.5-sm classifies on-site data and performs local work, and hands off material that needs further overall analysis to cip-5.5-im or cip-5.5-mm.

Calls allowed on the gateway

The gateway plan determines which calls each model allows. The table below shows the range opened by the LLM_FREE, TTS_FREE, and STT_FREE plans, which are applied automatically without a request. Models opened through other routes are not included.

cip-5.5-sm and cova-1 are not in this plan table. The cip-5.5-im-chat, cip-5.5-mm-h, cosa-tts, ivy-4-embedding, ivy-4-embedding-mm, and ivy-4-mm entries in the table are model IDs not covered in the product descriptions above. Which path each call key maps to is described in the Make API calls document.