Private AI chat on open-weights models
Candid AI. No fluff.
Most AI assistants are eager to please. They open with "Certainly!". They wrap every answer in em-dashes and headers. They tell you what they're about to tell you, then tell you, then summarise what they just told you. Berges AI doesn't.
Most AI chats
Certainly! That's a great question. Here's a breakdown of the key considerations — let me walk you through them step by step:
1. First Point — This is important because…
2. Second Point — Also worth noting…
Hope this helps! Let me know if you'd like me to expand on any of these.
Berges AI
Two things matter here. First, X, because Y. Second, Z. If you want it broken down further, say so.
Three layers. One conversation.
Most AI apps quietly hand your question to a lite model and pad the answer. Berges AI runs a frontier open-weights reasoning model on every turn, lets you pick which one, and cuts the padding.
Deep-thinking layer
A frontier open-weights reasoning model answers every question, and you watch it think. Pick DeepSeek, GLM, Qwen or Kimi from the sidebar. The models →
Pipeline layer
Pick a Berges pipeline: Reasoning, WebSearch, Brainstorm, or Empathy, to shape how the answer comes out. The pipelines →
Chatbot layer
Pick who answers: Luis is concise, Sara is friendly, Peter is technical. The chatbots →
Under all three, house rules keep every answer direct: answer first, no filler, prose first.
The same models and chatbots run in your terminal with the Berges CLI, an agent for your project. Included with Pro and Plus.
Built to extend you, not impersonate you
AI is great at some things. It's terrible at others. We've designed Berges AI to be honest about both.
AI is good at
- Drafting that first ugly version of anything
- Summarising long things into short things
- Finding patterns in noise
- Translating between languages, formats, and tones
- Suggesting the next reasonable step
You are still better at
- Knowing what actually matters to you
- Judging what's worth doing in the first place
- Reading the people in the room
- Living your life
- Taking responsibility
Your conversations stay yours
No conversation is stored until you save it. Saved conversations are encrypted at rest with AES-256. We don't read them, we don't sell them, we don't share them, we don't use them for training.
Encrypted at rest
Save a conversation and its messages and title are stored as AES-256 ciphertext, not plain text.
Never training data
We don't read your convos and we don't use them to train any model.
Delete anytime
Delete a saved conversation and confirm. Deleting removes it from our database right away. It drops out of our backups within 30 days, and conversation text stays encrypted inside them.
Open models. Open mind.
We run only open-weights models. Each one's license is on its model page.
Things people ask
Why three layers?
Most AI apps quietly hand your question to a lite model and pad the answer. Berges AI runs a frontier open-weights reasoning model on every turn, and two more layers sit on top: a pipeline you pick (Reasoning, WebSearch, Brainstorm or Empathy) that shapes how the answer comes out, and a chatbot you pick that sets the tone: Luis is concise, Sara is friendly, Peter is technical. Under all three, house rules keep it direct: answer first, no filler, prose first, today's date on hand, pasted links read.
Do you train on my conversations?
No. A conversation is stored only when you save it, encrypted at rest (AES-256), and it is never used to train any model, ours or anyone else's.
What happens when I run out of messages?
You can still read your conversations. On a paid plan, new messages pause until your monthly limit resets, or you can move to a higher plan. Each plan also has a short-term cap over a rolling window of a few hours, so a heavy burst can pause you briefly even with monthly allowance left; the usage page shows when it resets. The free trial is a one-time allowance that does not reset: once it is used, pick a plan to keep chatting.
Why open-weights models?
Three reasons: transparency (anyone can audit them), pricing (we can host them ourselves when it makes sense), and longevity (no vendor can pull them away).
Who makes Berges AI?
Berges Institute, a language school founded in NYC in 2013 and headquartered in Madrid, Spain. We started experimenting with large language models in October 2022, two months before ChatGPT launched. Our classes are still taught by human instructors, and always will be. Nothing from our classes or students is ever used to train a model. Berges AI is a separate product made based on our own research and experience.
What happened to Luis, Sara and Peter from Deep Spanish?
They live on in Berges AI. Deep Spanish was our first AI experiment, launched in October 2022: three Spanish practice chatbots built on an early language model. Luis was the sarcastic one, Sara the super-friendly one, and Peter the grammar nerd. Today they are the chatbot layer of Berges AI. They run on current open-weights reasoning models, answer in any language and on any topic, and each sets a different tone: Luis is concise, Sara is friendly, Peter is technical. Pick one from the sidebar.
Can I choose which model answers?
Yes. The sidebar lists four open-weights reasoning models: DeepSeek, GLM, Qwen and Kimi. DeepSeek is the quickest; the larger ones go deeper on the hardest questions and cost more units per reply. You can change it any time, and when you are signed in your choice is remembered across sessions. Your pipeline and chatbot still apply.
Can I paste a link?
Yes. Put a URL in your message and our server fetches the page and hands its main text to the model with your question, on any pipeline. Up to three links per message work, and the model reads the first 12,000 characters of each. The page stays available for follow-ups in the same conversation. Pages behind a login or a paywall will not load, and neither will PDFs or pages that need JavaScript. It is also the easy way past the message length limit.
When do answers use formatting?
Only when the content has that shape. Berges AI writes in prose by default, because formatting is where most AI padding hides. Steps come as a numbered list, code comes in a code block you can copy, and a comparison across the same points can come as a table. Short answers stay as plain sentences, and headings appear only in long answers with distinct sections. If you want a specific format, ask for it.
How do usage limits work?
Plans are measured in units, not messages, and larger models use more units per reply. Each plan has a monthly allowance and a short-term cap over a rolling window of a few hours, so a burst of heavy use can pause you briefly. The usage page shows both as percentages with the time to the next reset.
Does Berges AI search the web?
Yes, with the WebSearch pipeline selected. The model decides per question whether to search or read a page, and every search and fetch shows as a badge while the reply comes in, so you know when it went to the web. Pick Reasoning, Brainstorm or Empathy for no web search. Pasting a link works on any pipeline and fetches just that page.
Is there a terminal app?
Yes. The Berges CLI runs Berges AI as an agent in your terminal on macOS and Linux. It reads and edits files and runs commands in your project, on the same models and chatbots as the chat, and the same pipelines except WebSearch. It comes with Pro and Plus and uses the same allowance as the chat. Install it with Homebrew: brew install bergesinstitute/tap/berges.
Made by a Spanish language school
Berges AI is built by Berges Institute, an online Spanish language school. Language is our craft, so how a machine handles it is not a side interest for us.
Early, on purpose
We started building with large language models in October 2022, two months before ChatGPT launched. Berges AI is what grew out of that work.
Taught by people
Our classes are run by human instructors, the way they have been since day one and the way they'll stay. Berges AI is a separate product, kept separate on purpose.