Meta AI Gives Norwegian and Sami the World's Best Language-Model Embeddings

That time is now over.

What is F2LLM-v2?

F2LLM-v2 is a family of 8 general-purpose multilingual embedding models, ranging from 80 million to 14 billion parameters, trained on 60 million carefully curated data points.

Language coverage: 200+ languages, with a particular focus on previously underrepresented medium- and low-resource languages.

Norwegian and Sami are now first-class citizens in the world of AI.

Performance That Beats Everything

F2LLM-v2-14B (the largest model) ranks #1 on 11 MTEB benchmarks – the industry standards for measuring embedding quality.

But what's impressive is not just the largest model. The entire family sets new standards:

| Model size | Use case | Performance |

|-----------------|-------------|--------|

| F2LLM-v2-80M | Mobile/Edge | State-of-the-art for resource-constrained devices |

| F2LLM-v2-14B | Server/Cloud | #1 on 11 MTEB benchmarks |

Wherever you run AI – from phone to cloud service – there is now an F2LLM-v2 model optimised for your use case.

Technical Innovations

Meta combines three powerful techniques:

1. Matryoshka Learning

Like a Russian matryoshka doll, the model contains several "layers" of representations. You can use 256, 512, or 1024 dimensions depending on how much memory you have available.

2. Knowledge Distillation

Smaller models learn from larger models, maintaining quality at a fraction of the size.

3. Model Pruning

Removes redundant "weight" from the model without losing performance.

The result: Maximum efficiency with no compromise on quality.

What Does This Mean for Norway?

1. Norwegian-Language AI Finally Becomes Competitive

Before: English embeddings + hope that translation works

Now: Native Norwegian processing at a world-class level

Projects such as nettsak.no, openinfo.no, and InfoDesk can now use embeddings that understand Norwegian context just as well as English models understand English.

2. Sami Language Revitalisation Through AI

For the first time, the Sami languages (Northern Sami, Southern Sami, Lule Sami) have access to SOTA AI embeddings.

Implications:

• Automatic translation Sami ↔ Norwegian on a par with Google Translate English ↔ Spanish
• Sami voice assistants and chatbots
• Digitisation of Sami cultural heritage with AI-driven search tools

3. Norwegian Start-ups Gain a Competitive Edge

With F2LLM-v2, Norwegian AI companies can:

• Build search engines that understand Norwegian idiom
• Create chatbots that don't sound like a bad Google Translate
• Develop semantic search for legal documents (the Norwegian legal context)

Open-Source Everything

Meta is following its new approach to AI: Full transparency.

• ✅ All 8 models (open weights)
• ✅ Training data (60M samples)
• ✅ Source code
• ✅ Intermediate checkpoints

Everything is available for free on Hugging Face.

Why Is Meta Doing This?

Strategic gain: The more people build on Meta's models, the more data and feedback Meta receives. Open source is their competitive advantage against OpenAI and Anthropic.

Ethical gain: Democratisation of AI. Small languages should not be second-class citizens in the AI revolution.

Technical gain: Open source means thousands of developers who find bugs, suggest improvements, and build applications Meta would never have thought of.

What Happens Now?

F2LLM-v2 is already available. Norwegian developers can:

1. Download the model from Hugging Face (search "F2LLM-v2")
2. Integrate it into existing systems (HuggingFace Transformers compatible)
3. Fine-tune on Norwegian datasets for even better performance

The first generation of Norwegian AI applications based on F2LLM-v2 will probably launch within weeks.

We are no longer talking about "when Norwegian AI will be good."

That time is now.

Sources:

• ArXiv: 2603.19223v1 (Zhang et al., Meta AI)
• Meta AI Research Blog
• Hugging Face Model Hub

Published: 21 March 2026