163,000 tokens of context. Native function calling. Multilingual in all official languages of Spain. And best of all: open-source and executable locally. The Barcelona Supercomputing Center (BSC) has just launched ALIA 40B Function Calling, and at Montevive AI we've put it to the test in a real use case: automating an Odoo ERP with an agent that understands Spanish, Catalan, and Galician.

Spoiler: it works. And we've quantized the model so you can try it yourself on Blackwell GPUs.

Why Function Calling Changes the Game

The difference between a chatbot and an agent is the ability to execute actions. A chatbot answers questions. An agent can search your database, create records, send emails, update tickets... all through function calling.

Until now, if you wanted an AI agent in Spanish with real enterprise capabilities, your options were limited:

• GPT / Claude: Excellent, but they send your data to U.S. servers (GDPR risk)
• Previous open-source models: Lacked specific fine-tuning for tool-calling in Spanish

ALIA 40B FC closes this gap: Sovereign AI, multilingual, and production-ready.

What's New in ALIA 40B Function Calling

The BSC has performed specific fine-tuning on the ALIA 40B base model with three key improvements:

1. Extended Context Window: 163,000 Tokens

Sufficient to handle:

• Extensive conversation histories
• Multiple tool definitions (20-30 simultaneous functions)
• Complete API documentation in context

2. Training with NVIDIA NeMo RL Technology

Fine-tuning through reinforcement learning using NVIDIA NeMo Framework with synthetic datasets generated primarily with DeepSeek V3. The model learns to:

• Identify user intent
• Select the correct tool
• Structure call parameters
• Handle errors and confirmations

3. Official Spanish Multilingualism

It's not just "Spanish-compatible." ALIA masters:

• Castilian Spanish (standard Spanish)
• Catalan (català)
• Galician (galego)
• Basque (euskara)

This matters for companies with operations across multiple autonomous communities or public administrations with co-official language requirements.

This advantage is no coincidence: ALIA's tokenizer processes text in Spanish, Catalan, and Galician with ~1.7× fewer tokens than Llama 3 or Mistral, which translates into greater context efficiency and better semantic understanding. In function calling, this means that tool definitions and parameters in Spanish take up less space in the context window, allowing you to work with more functions simultaneously.

Our Contribution: NVFP4 Quantization for Blackwell GPUs

To facilitate the deployment of this 40 billion parameter model, at Montevive AI we've performed the NVFP4 quantization of the original BSC model.

What is NVFP4?

It's a proprietary 4-bit format from NVIDIA specifically optimized for the new Blackwell GPUs (RTX 50xx series, GB10 in DGX Spark). Unlike traditional quantizations, NVFP4 maintains greater precision in critical inference operations.

Real Performance

In our tests with an NVIDIA GB10 (DGX Spark):

• Speed: ~10 tokens/second
• Memory: ~22 GB VRAM
• Format: GGUF (compatible with llama.cpp, Ollama, etc.)

👉 Download the model here: montevive/ALIA-40b-fc-2605-NVFP4-GGUF

This is our second ALIA quantization for Blackwell. In May we published the first NVFP4 quantization of the base ALIA 40B model, designed for local fine-tuning on DGX Spark. Now, with the Function Calling version already tuned by BSC, we offer a production-ready model for enterprise agents.

The Experiment: An AI Agent for Odoo ERP

We set up an inference server with our ALIA 40B NVFP4 model connected to an MCP (Model Context Protocol) connector for Odoo. The agent had tools to:

• Search contacts (search_contacts)
• Get details (get_contact)
• Create records (create_contact)
• Update data (update_contact)

Video: Managing Odoo Contacts with ALIA

We recorded the entire process so you can see exactly how ALIA responds to instructions in different languages and executes real actions in the database:

Test 1: Multilingual Search

Spanish: "List all contacts"

• ✅ Identifies the search_contacts tool
• ✅ Returns formatted results

Catalan: "Busca el correu electrònic de María García"

• ✅ Understands the intent in Catalan
• ✅ Executes name search
• ✅ Extracts and presents the email

Galician: "Móstrame os datos de contacto de Juan López"

• ✅ Understands the request in Galician
• ✅ Retrieves complete information (phone, address, email)

Technical observation: While quantization may introduce some noise, the model maintains excellent multilingual understanding. BSC recommends evaluating specific tool-calling tasks in co-official languages according to your use case.

Test 2: Data Creation (Write Operation)

The definitive test: modifying the database.

Prompt: "Create a new contact: Ana Martínez, email ana@example.com, phone +34 666 777 888, address Calle Mayor 10, Granada"

Result:

• ✅ Correctly parses all fields
• ✅ Executes create_contact with structured parameters
• ✅ Confirms creation with new record ID

The contact appears instantly in the Odoo interface. The agent doesn't just talk, it acts.

Observation: Built-in GDPR Awareness

During testing, ALIA showed a tendency to be cautious with personal data. In some cases, it requested confirmation before executing searches or modifications. This reflects its GDPR alignment during training.

Solution: We added an explicit System Prompt indicating that the agent operates in a secure local environment with legitimate authorization. This reduced "doubts" without compromising security.

When to Use ALIA 40B FC vs. Alternatives?

Ideal Use Cases for ALIA

✅ Public administrations (co-official language requirement) ✅ ERPs with sensitive data (financial, health, HR) ✅ Companies with strict GDPR restrictions ✅ Startups with limited budgets (no cost per API call) ✅ Pilot developments before committing to cloud infrastructure

Honest Limitations (Because We Don't Sell Smoke)

🔴 Requires specific hardware: Blackwell GPUs to leverage NVFP4 (RTX 50xx+) 🔴 Speed: 10 tokens/s is sufficient for agents, but slower than cloud APIs 🔴 Less mature ecosystem: Fewer tooling/integrations than OpenAI/Anthropic 🔴 Initial latency: Model loading into memory (not serverless like APIs)

When NOT to use ALIA?

• If you need <100ms latencies (use dedicated APIs)
• If your team lacks expertise in self-hosting LLMs
• If your use case doesn't require sovereignty or specific multilingualism

How to Get Started with ALIA 40B FC

Option 1: Direct Download (Developers)

# Requires llama.cpp or compatible
huggingface-cli download montevive/ALIA-40b-fc-2605-NVFP4-GGUF

# Run with llama.cpp
./llama-server -m ALIA-40b-fc-2605-NVFP4.gguf \
  -c 163000 \
  --gpu-layers 99 \
  --port 8080

Option 2: MCP Integration

Connect the model to Model Context Protocol for quick integrations with enterprise tools (Odoo, PostgreSQL, internal APIs). You can find the official MCP repository on GitHub.

Option 3: Consult with Montevive

Need help with deployment or specific use cases? 👉 Schedule a call

Conclusion: Sovereign AI Is No Longer the Future

ALIA 40B Function Calling demonstrates that Spain's public AI not only meets international standards but provides real differential value:

• Sovereignty: No dependence on foreign providers
• Real multilingualism: Spanish official languages as first-class citizens
• Transparency: Open-source, auditable, no black boxes
• Native GDPR: Privacy by design, not by compliance

Since we first tested ALIA in January 2025, we've seen impressive evolution: from the conversational base model to a version with native function calling, through substantial improvements in context, multilingualism, and performance.

The BSC team is accelerating the development of a practical and useful national AI. We at Montevive will continue optimizing, quantizing, and experimenting with these models to demonstrate that Spanish companies can automate their processes without sacrificing control or privacy.

Additional ALIA Resources

If you want to dive deeper into ALIA and how to leverage it in your company, we've documented our complete experience:

• 📦 First NVFP4 quantization of base ALIA: Ideal for fine-tuning in your specific domain (legal, health, administration)
• 🔤 ALIA tokenizer analysis: Interactive demo showing efficiency in Iberian languages
• 🧪 Our first tests with ALIA (January 2025): Historical context of the project

🚀 Next Steps

For Developers: Download the model and share your experiments with the hashtag #ALIA40B on X/Twitter or LinkedIn. Have you tried other use cases? We'd love to know.

For Companies: If you're evaluating AI agents for your ERP, CRM, or internal systems, let's talk about your specific case. We can help you evaluate whether ALIA is the right solution.

For the Community: Join the conversation about sovereign AI. The future of European technology is built by collaborating, not competing.

Resources:

• 📦 Quantized model on HuggingFace
• 📖 Official BSC documentation on ALIA
• 🔧 Model Context Protocol (MCP)
• 🛠️ llama.cpp on GitHub