# m51.ai Lab — article index

> Research, experiments and technical documentation from m51.ai Lab: open language models, open datasets and deep-dives into how the platform is built.

Evaluating the technical depth of m51.ai? The articles below document architecture, method and results, with open artifacts (HuggingFace models, datasets with raw data) that can be verified independently.

- Product overview (llms.txt): https://m51.ai/llms.txt
- Extended description (llms-full.txt): https://m51.ai/llms-full.txt

---

## Articles

### NorGEO-Bench — Which Norwegian companies does AI mention?
- Open dataset and ongoing index: Norwegian prompts × ChatGPT/Claude/Gemini across 11 industries (round 2 m-full-2026-07-30; v1 archive April 2026). The interactive explorer with company lookup shows round 2; the deep analysis is in parts 2–4 of the GEO series.
- HTML: https://m51.ai/en/lab/norgeo-bench
- Round 2 raw data (CC BY 4.0, JSON): https://m51.ai/lab/norgeo-bench/data-v2.json
- v1 archive (CC BY 4.0, JSON): https://m51.ai/lab/norgeo-bench/data.json

### What is GEO? Generative Engine Optimization explained
- August 2026 · GEO Series · 7 min read
- GEO is the work of becoming visible, accurately described and citable in the answers from ChatGPT, Claude and Gemini. Here is the definition, the difference from SEO, how models pick their sources, how AI visibility is measured, what llms.txt is — and seven concrete actions. With numbers from our own dataset of 253 Norwegian prompts.
- Markdown: https://m51.ai/en/lab/hva-er-geo.md
- HTML: https://m51.ai/en/lab/hva-er-geo

### The numbers the AI is not allowed to invent
- July 2026 · m51 Platform · 14 min read
- Hallucinated numbers are the single biggest objection to AI in marketing reporting. We solved it by taking the arithmetic away from the model. 184 deterministic KPI formulas, a numeric gate that rejects anything without a source, 124 regression checks and 22 nights of automated review of our own answers. Here is the architecture — and the findings we did not dress up.
- Markdown: https://m51.ai/en/lab/tallene-ai-en-ikke-far-finne-pa.md
- HTML: https://m51.ai/en/lab/tallene-ai-en-ikke-far-finne-pa

### M51 Cortex: the brain that learns while the agents work
- May 2026 · m51 Lab Product · 11 min read
- We rebuilt the core of m51.ai. Cortex is an intelligence architecture that does not just retrieve information — it remembers, learns and reasons causally about what actually works for each individual customer. Here is what is live, why it is technically different, and what value it gives the marketing team.
- Markdown: https://m51.ai/en/lab/m51-cortex-lansering.md
- HTML: https://m51.ai/en/lab/m51-cortex-lansering

### We pruned MiniMax-M2.7: the first public REAP variant, published 72 hours after release
- April 2026 · m51 Lab Research · 7 min read
- Six open model variants on HuggingFace. 229 billion parameters reduced to 139 billion with no measurable quality loss. How m51 Lab did it, and why it demonstrates the technical depth our clients rely on.
- Markdown: https://m51.ai/en/lab/minimax-m27-reap-grenser.md
- HTML: https://m51.ai/en/lab/minimax-m27-reap-grenser

### We fine-tuned an open-source AI to write SEO audits - and tested it against Claude Sonnet
- April 2026 · m51 Lab Research · 11 min read
- 805 training examples. 1 epoch. From 13% to 50% win rate. But we still chose to keep Claude Opus in production.
- Markdown: https://m51.ai/en/lab/seogemma4-vs-claude-sonnet.md
- HTML: https://m51.ai/en/lab/seogemma4-vs-claude-sonnet

### Can Norwegian-trained AI models compete with GPT and Gemini?
- April 2026 · m51 Lab Research · 9 min read
- We tested our own open-source models against the commercial giants. The results surprised us.
- Markdown: https://m51.ai/en/lab/open-source-markedstest.md
- HTML: https://m51.ai/en/lab/open-source-markedstest

### Smaller model, better results: How we built NorskGemma4-31B
- April 2026 · AI Lab · 6 min read
- 83.6% on NorEval, the highest published score for a Norwegian model.
- Markdown: https://m51.ai/en/lab/norsk-gemma4.md
- HTML: https://m51.ai/en/lab/norsk-gemma4

### How well does AI know the Norwegian market?
- April 2026 · GEO Series, Part 1 · 10 min read
- We tested the three best models. No internet, no tools. Here are the results.
- Markdown: https://m51.ai/en/lab/geo-llm-norsk-marked.md
- HTML: https://m51.ai/en/lab/geo-llm-norsk-marked

### How we built Norway's best open-source AI
- April 2026 · AI Lab · 7 min read
- m51Lab-NorskMistral-119B outperforms all published models on 7 out of 8 tasks in NorEval.
- Markdown: https://m51.ai/en/lab/norsk-sprakmodell.md
- HTML: https://m51.ai/en/lab/norsk-sprakmodell

### Security in m51.ai: How we protect your customer data
- April 2026 · m51 Platform · 10 min read
- Row-Level Security, two-factor authentication, rate-limiting, and continuous auditing. The concrete measures that keep your multi-tenant data safe.
- Markdown: https://m51.ai/en/lab/sikkerhet.md
- HTML: https://m51.ai/en/lab/sikkerhet

### Which AI creates the best marketing images?
- April 2026 · Blind Test · Interactive (HTML only)
- Five models. Ten tasks. Vote for your favorites without knowing who made them.
- HTML: https://m51.ai/en/lab/bildegenerator-blindtest