Aunoa has introduced AunoaLLM , its own platform for specialising and deploying language models through adapters. On October 6, 2026, the Valencian company announced the completion of the Adaptadores research project from which this technology emerged. Its proposition uses open models as a base and adds specialist knowledge without repeating full training for every application domain.
The announcement sits within a specific business: conversations between companies and their customers. Aunoa markets agents that answer ecommerce enquiries, capture contacts from digital campaigns, qualify opportunities and pass information to commercial systems. The project validated five demonstrators, including ecommerce. AunoaLLM adds a model-development and management layer to this strategy of turning conversation into a tool for service and sales.
The reported technical results make the development more concrete. The platform has been tested with four base models, ranging from seven billion to more than thirty billion parameters. The company reports switching between specialist adapters within twelve to twenty seconds when the base model is already loaded into memory. Its announcement also describes training, evaluation, retrieval of proprietary information, deployment and traceability within one environment.
One common base and several specialisations
The project is called Adaptadores y Técnicas Complementarias en la utilización de modelos de lenguaje. Aunoa presents it as a research, development and innovation initiative intended to strengthen its own generative-AI capabilities. The stated aim is greater control over data, privacy and deployment economics, with less dependence on external services. Infrastructure and knowledge specialisation consequently become part of the company's product-development work, alongside the conversational functions offered to business customers.
The central idea separates a base model from its adaptation to a domain. AunoaLLM uses LoRA and QLoRA, techniques that work with adapters instead of retraining every model parameter. In the company's explanation, LoRA freezes the base and trains small additional matrices; QLoRA combines that adaptation with a base quantised to four bits. The intended result is to create specialisations by sector or customer with lower requirements than full training would involve.
That distinction matters to commercial conversational AI. A question about delivery, a financial-service request and a query about internal company knowledge require different information and vocabulary. The project proposes adapting a base to those domains through specialist components. This separates development tasks: the company can manage which adaptation is used, evaluate it and deploy it inside the platform while retaining an identifiable relationship with the underlying model from which it started.
What was tested and what twelve to twenty seconds means
Aunoa identifies four models in its validation: Phi-4 14B, Qwen2.5-7B, Qwen2.5-32B and Gemma 4-31B. The catalogue contains different nominal sizes and two variants within the same family. According to the announcement, working with models above thirty billion parameters extended the capabilities initially envisaged for the project. This describes the technical scope the company reports having reached, accompanied by specific model names and sizes.
Base models reported in AunoaLLM validation
Nominal sizes of the four models identified by Aunoa in its October 6, 2026 announcement of results from the Adaptadores project.
billion parameters
Tap or hover for details. All values are also available in the data table.
Collapse notes
Validation catalogue reported by Aunoa. Nominal size is not a score for quality, response speed or commercial performance.
Explore the data table
| Category | Nominal size |
|---|---|
| Phi-4 14B | 14 billion parameters |
| Qwen2.5-7B | 7 billion parameters |
| Qwen2.5-32B | 32 billion parameters |
| Gemma 4-31B | 31 billion parameters |
The twelve-to-twenty-second interval concerns switching adapters while the base model is loaded. It measures an operation performed by the platform managing specialisations, rather than the time a shopper waits for a chat response. The condition matters: loading the base and changing an adapter are different operations. The figure explains why Aunoa relates the function to deploying several assistants over shared infrastructure, with a separate specialist component for each knowledge domain.
The five demonstration domains are financial customer service, telecommunications support, ecommerce, public services and queries about regulations or corporate knowledge. Their breadth reflects the adaptation problem: adjusting one base to different knowledge and conversations. For this story, ecommerce supplies a direct connection with the digital relationship between a brand, its catalogue and the people asking questions or making purchases, within the business context of Aunoa's conversational offer.
The results announcement concerns completion of the project and development of this platform. Aunoa's current commercial pages help explain where that technology sits within its activity: acquisition, service, connections with business tools and continuity across conversations. These are two related levels of the company's work, developing specialist models and offering conversational functions to business customers. Keeping those levels identifiable makes the product-development decision easier to understand.
Training is only one part of the journey
AunoaLLM is presented as a platform for managing the full adapter lifecycle. It includes a training and fine-tuning engine, a graphical configuration and evaluation interface, and deployment and administration services. The proposition brings together tasks performed before and after preparing a specialisation. For a developer of assistants, this organises creation of the component and its subsequent use inside the same working environment, rather than treating training as the complete product.
Information retrieval, or RAG, is another listed function. Its described role is to answer using the organisation's own knowledge. The combination distinguishes adapting a model from supplying relevant information for a query. In a commercial setting, a response can require business-specific knowledge while the assistant must interpret the question and maintain a useful customer conversation. The platform description therefore addresses both specialised behaviour and access to information.
The platform adds version history, behaviour monitoring and an inference module for testing and validation before deployment. It also includes comparison between models and versions. This places evaluation inside the development process alongside training. Producing a variant is one step; the environment is designed to examine it, preserve its history and support a decision about which version to use. Those functions give operational continuity to the process of preparing an assistant.
The functions provide a business reading of the technical announcement. Aunoa is bringing decisions about specialisation, knowledge, versions and deployment into its own platform. For the company, the value lies in turning applied research into an asset that can evolve and reach the market. The project extends its development tools, while its product pages show the commercial problems at which it directs conversational activity and the systems with which that activity needs to connect.
From a digital advertisement to a contact in the CRM
The current lead-acquisition page describes a journey starting before contact with a salesperson. An agent talks to the visitor, collects information, profiles the opportunity and sends it to the CRM. Aunoa proposes questions adapted to the acquisition process and the brand's tone. Conversation forms part of a marketing and sales operation: it turns an interaction into information a commercial team can use to continue following up, linking the initial digital encounter with subsequent work.
The offer identifies several entry points. One is a campaign landing page; another is a conversation originating from Instagram Ads, Meta Click-to-WhatsApp or Google Ads. It also describes webchat activated by behaviour such as pages visited, scrolling or time spent on the site. Channels and moments differ, but the proposed journey brings conversation, identification of interest and delivery of information to the sales system together within the acquisition process.
Lead scoring adds classification. The page discusses assessing interest, budget, profile and behaviour, with criteria adapted to the business. It also mentions offering an appointment with the sales team and managing availability and preferences. These are functions Aunoa markets for organising the transition between digital conversation and human involvement. The commercial team can receive a richer context than isolated contact details, including the information gathered while the visitor explains an interest or need.
Information can go to the CRM or a marketing platform, and the page includes email or Slack alerts. Tools shown in the offer include Salesforce, HubSpot, Zendesk and Freshdesk alongside ecommerce platforms. For the acquisition mechanism, the destination of the data matters: conversation connects with the subsequent work of a service or sales team. This is how the proposition links a dialogue in a digital channel with the systems used to manage commercial activity.
The journey places model specialisation within Aunoa's business. Conversations handle sector vocabulary, qualification criteria and knowledge of an offer. AunoaLLM addresses adaptation to domains; the commercial page explains the tasks the company proposes to solve through its agents. Their relationship can be understood through these requirements for knowledge and continuity. The technical platform and the customer-facing workflow describe different levels of the company's conversational product strategy.
In a store, conversation continues after the purchase
The ecommerce offer covers questions before purchase and subsequent service. The page describes enquiries about stock, sizes and features, payment methods, shipping costs and returns. It also proposes alternatives when products are unavailable and helping customers start a purchase from the conversation. These are moments when catalogue details and store policies become information a user needs to progress, bringing customer assistance close to the practical work of selling online.
After purchase, Aunoa presents order-status enquiries, address changes, cancellations and data updates. It mentions connections with Shopify, PrestaShop and Magento and support tools. The marketed capability goes beyond a general answer: an agent needs to relate the enquiry to information about the transaction and, for some requests, to an action within the store's system. That separates discussing a policy from processing a particular customer's request using the relevant operational information.
Recording the interaction is also part of the proposition. The page says conversations can be summarised and passed to the CRM or corresponding ticket. It adds incident classification and assignment to the appropriate department. History and routing connect automated service with team activity: a person can receive structured information about what has already happened in the conversation, giving the handover a context and a destination inside the business's service process.
The frequently asked questions distinguish connections with established commercial tools from integration with proprietary systems. Aunoa lists connectors for Shopify, PrestaShop, Magento, Zendesk, Salesforce, HubSpot, Freshdesk and Jira, and proposes a REST API when customers use their own software. This makes the intended connection between conversation and business data more concrete. A catalogue, order and incident can reside in different systems; the offer describes connection points through which they can relate to the assistant.
The offer includes a hybrid model, with complex enquiries passed to human agents, and conversation across channels such as WhatsApp, webchat and Instagram. That continuity is relevant to a brand handling product questions, incidents and later purchases from the same customer. In the commercial pages, automation and human attention appear as parts of a coordinated service, sharing information while performing different functions in the relationship between company and customer.
The company also describes proactive training, supervision of resolution metrics and adjustments during a project's life. This operational service accompanies the assistant after launch. AunoaLLM, meanwhile, includes versions and evaluation at model level. Both descriptions place evolution and monitoring within Aunoa's technology offer, rather than reducing the product to installing a chat interface on a website. The distinction connects development capabilities with the ongoing service through which the company markets conversational automation.
In its commercial answers about support, Aunoa proposes an agreed automation threshold and action when the rate falls below that level, without waiting for the customer to raise an incident. It presents support as part of the contract. That detail defines an operational responsibility within the service offered: observing the assistant's evolution and adjusting its functioning. The proposition combines initial integration, hybrid attention and later monitoring within a continuing relationship with the business using the service.
Spanish context: AI alongside other data and management layers
The business reference from the Instituto Nacional de Estadística helps locate the adoption environment. In the first quarter of 2025, 21.1% of internet-connected enterprises with ten or more employees in the sectors covered by its survey used artificial intelligence. Within the same base, cloud-service use was 44.3%, data analytics 41.4% and CRM 34.5%. These are indicators of different tools that a company can use simultaneously, rather than competing categories that must add up to one hundred.
AI and data-tool use in Spanish enterprises
First quarter of 2025: percentages of internet-connected enterprises with ten or more employees in the sectors included in the INE ETICCE survey.
% of internet-connected enterprises
Tap or hover for details. All values are also available in the data table.
Collapse notes
Historical Spanish context, not Aunoa results or customers. Indicators are compatible: one enterprise can use several tools.
Explore the data table
| Category | Enterprises using each tool |
|---|---|
| Artificial intelligence | 21.1 % of internet-connected enterprises |
| Cloud services | 44.3 % of internet-connected enterprises |
| Data analytics | 41.4 % of internet-connected enterprises |
| CRM | 34.5 % of internet-connected enterprises |
That combination is relevant to an offer connecting agents, proprietary knowledge and commercial systems. Adding AI to a conversation does not remove the other layers of work: product information, customer records and sales follow-up retain different functions. Aunoa's proposition operates between them when it discusses RAG, CRM connections, ecommerce transactions and transfer of cases to people, linking the conversational interface with the information and responsibilities behind it.
The percentages concern Spain and the first quarter of 2025 as historical context, rather than Aunoa customers or its October launch. They describe the coexistence of technologies within a defined enterprise population. The recent event is the outcome of the Adaptadores project and presentation of a platform through which the company proposes to develop its own specialisations for different areas of activity, including the ecommerce demonstrator identified in its announcement.
Applied research and transfer to the market
The project was developed within the INNOVATeiC-CV programme under IVACE+i's INNOVA-CV 2025 call. According to Aunoa, it is funded by the Generalitat Valenciana with European funds. The call provides the project's framework; the announcement of completion and the results giving rise to AunoaLLM came on October 6, 2026. The company places transfer of this knowledge to the market among its next steps, connecting research with future development of products and services.
The development brings two decisions together: having its own platform for working with open models, and organising specialisation as a cycle of training, evaluation, versions and deployment. In its commercial activity, Aunoa directs agents towards acquisition and service conversations. The business connection lies in adapting knowledge and operation to those uses, with development infrastructure the company presents as part of its technological autonomy and its capacity to create specialist conversational solutions.
Who's who and how to read the figures
Aunoa developed and presented AunoaLLM. Its terms of use identify Aunoa Software S.L., registered with Valencia's Commercial Registry. AunoaLLM is the platform emerging from the Adaptadores project; the lead-acquisition and ecommerce pages discussed here belong to the commercial offer observed in October. The principal announcement dates from October 6, 2026, while the functions marketed on those pages provide context for the company's activity in conversations, service and sales.
Parameters express the nominal size of the base models identified by the company. The four names in the validated catalogue do not constitute a quality ranking. The twelve-to-twenty-second interval concerns adapter switching with the base in memory. LoRA and QLoRA are the adaptation techniques used; RAG is information retrieval to answer with proprietary knowledge. These terms describe different components of the platform and the role each performs in preparing a specialist model.
Lead scoring classifies opportunities through commercial criteria; a CRM organises customer relationships and follow-up. The Instituto Nacional de Estadística supplies Spanish adoption context using a base of internet-connected enterprises with at least ten employees in covered sectors. IVACE+i and the Generalitat Valenciana provide the research-support framework described by the company. Each set of figures retains its role: reported technical scope, enterprise environment and commercial development.
In this article
Marian Vasile
Marian Vasile is a MarTech and growth professional. His work spans product marketing, SEO, paid advertising and deep online tracking, including behavioral analytics and attribution.
