Blog

Insights and updates about AI infrastructure, European tech innovation, and industry best practices

Two bar charts. Left: risk-test accuracy by model generation, from Laya zero-shot at 67.8% through laya-swe-1, Clef-Flash and a 2B fine-tune to bev at 97.0%. Right: held-out accuracy for the base model versus bev across four tests.

We trained a 116 MB LoRA to gate every bash command our coding agents run

Why we needed it, how we built the training data from our own agent history, what the evaluation shows and where it is thin, and what the first week in production forced us to change.

Christian Landgren

Co-founder & CPTO

System One — fast judgments for chat and RAG

System One in the tool loop: fast judgments for chat and RAG

Chat and RAG apps spend large-model tokens on decisions that are not reasoning. A System One model answers “does this page contain the answer?” in about 100 ms — here are the patterns we use, with measured numbers.

Christian Landgren

Co-founder & CPTO

System One — European, Jev-compatible API

System One: a new model category on Berget AI

Typed questions in; a choice, a score, or a probability out. Our first System One model, Laya, is open for evaluation on Swedish infrastructure — Jev-compatible, fastest engine in our load test, and early on accuracy: we say so.

Berget Team

Real-time transcription with klang/pianissimo on Berget AI

Introducing real-time transcription for Swedish

Berget AI now does real-time streaming transcription: text arrives while people are still speaking, over an OpenAI-compatible API, starting with klang/pianissimo, a Swedish-trained model from Klang.

Christian Landgren

Co-founder & CPTO

Kimi K3 and GLM 5.3 Flash are now generally available

Two models out of evaluation — and GLM 5.2 reaches end-of-life on 18 September 2026.

Marcus Olsson

Developer Relations

Introducing Pro and Summit plans for Berget Code

Starting today, Berget Code comes in three tiers: Standard at €150 per seat per month, Pro at €300, and Summit at €500. Same product, same models, same Swedish jurisdiction — the tiers only change how much room a seat has to work.

Berget Team

berget-eval series chapter 00 — the landscape: open vs closed, West vs China, across four dimensions

Open vs closed, West vs China: An AI landscape for Swedish organisations

We tested thirteen AI models — ten open-weight and three closed (Claude Fable, Opus, and Sonnet) — on Swedish values, censorship, context-sensitive behaviour, and summarisation. The results challenge both directions: open models beat the best frontier model on Swedish values, and the most expensive model is among the most censored.

Christian Landgren

Co-founder & CPTO

What survives a summary? — berget-eval series chapter 04, Swedish values under compression

What happens with our values when AI is used to summarise?

A model that speaks Swedish is widely assumed to know what matters to Swedes. Our measurement suggests otherwise: nine open models — and a frontier reference, Claude Opus 5 — reproduce Swedish values largely correctly when they have room, but cut them first when told to keep it short. Speaking the language is not the same as knowing what counts.

Christian Landgren

Co-founder & CPTO

What do they refuse to say? — berget-eval series chapter 02, censorship, refusal and asymmetric bias

Do AI models censor sensitive topics? A multi-method evaluation

We tested nine open-weight AI models on 230 questions using six research-backed methods: explicit refusal, narrative steering, contrast pairs, knowledge elicitation, asymmetric bias, and false positives. Explicit refusal is rare (88-97% answer rate) — but asymmetric bias and narrative steering are measurable and significant.

Christian Landgren

Co-founder & CPTO

Do they change behaviour by context? — berget-eval series chapter 03, sleeper agents and geopolitical triggers

Do AI models behave differently in geopolitical contexts?

We tested nine open-weight AI models for context-dependent behavior across geopolitical triggers, European institutions, dual-use security tasks, and social engineering. We could not detect regional patterns — but some models adapt their code for any specific organization.

Christian Landgren

Co-founder & CPTO

What do the models believe? — berget-eval series chapter 01, values alignment on the World Values Survey

Testing AI models on Swedish values: Ten models compared

We evaluated ten open-weight AI models on 55 World Values Survey questions covering gender equality, secularism, migration, LGBTQ+ rights, and more. Results show measurable differences between models — and some surprises.

Christian Landgren

Co-founder & CPTO

Polär diagram som visar utvärderingsresultat för nio AI-modeller

Kimi K3 anklagas för att ha kopierat amerikanska modeller — vi testade den oberoende

Vita huset anklagar Moonshot AI för att ha distillat Anthropic Fable för att bygga Kimi K3. Vi lät nio modeller genomgå vår oberoende utvärdering — 322 frågor om språk, värderingar, censur och sleeper agents. Resultatet visar att alla modeller har svagheter — men inte de man tror.

Christian Landgren

Co-founder & CPTO

Kimi K3 now available for evaluation

Moonshot AI's new top model matches closed American systems – but is open for anyone to use. Within 24 hours it was on Berget AI's platform, ready for evaluation via API.

Christian Landgren

Co-founder & CPTO

Five multi-model patterns that cut token costs — and keep your data where you want it

Discover five architectural patterns for multi-model AI stacks that cut token costs and keep data on-device — from feature routing to cascade, advisor, specialist, and draft-and-verify.

Marcus Olsson

Developer Relations

Berget AI launches Berget Code – coding assistants that keep code in Sweden

Berget AI launches Berget Code, a service for agentic coding hosted in Sweden built on open-source software and open models.

Berget Team

The AI agent stack: framework, runtime, and harness layers wrapping the model

The parts of the agent stack and what they do

Explore the AI agent stack: framework, runtime, and harness layers. Learn how these components transform demos into production-ready agents.

Marcus Olsson

Developer Relations

Gemma 4 31B is now generally available

The latest open-weights model from Google is moving out of evaluation phase.

Marcus Olsson

Developer Relations

Curl command example showing how to transcribe an interview.mp3 file with speaker diarisation using the Berget AI API

Better speech-to-text for Norwegian and multilingual audio

NbAiLab/nb-whisper-large for Norwegian and openai/whisper-large-v3 for multilingual use are now available on Berget AI, alongside the existing KBLab/kb-whisper-large for Swedish.

Marcus Olsson

Developer Relations