Every model carries an origin label. Downloads are from HuggingFace.
Default American pick. Google's official quantized release. Native vision. Best for 24GB cards.
Alternative pick. Flag origin in security reviews.
Mid-size. Good balance for 12-16GB cards.
Smaller Gemma. Fits 8-12GB cards.
Smallest. Runs on modest GPUs.
Start a conversation, or choose a preset below.
Run renewal reminders and a new-member welcome from approved segment facts.
Coordinate a chapter meeting without turning proposals into assignments.
Draft conference copy around confirmed sessions and approved prices.
Report CE activity without posting credits the records do not support.
Separate what a committee decided from what was only proposed.
Draft outreach from confirmed packages; pending inventory stays a note.
Turn aggregate membership figures into a snapshot leadership can trust.
Catch leading and double-barreled survey questions before they field.
Query the site knowledge graph through any model.
Prepare a dataset from chats; generate an Unsloth training script.
Plan a GPU fit-check and harness install before downloading.
No presets match that search.
The entire conversation counts toward context.
Use public or synthetic information. Model downloads require Load.
Conversations save on this device. Manage this in Settings.
Before you start: The three boxes below are optional and start empty. Nothing is pre-filled. Fill in any of them with your own information and the AI will use what you write to give you relevant answers; leave them blank for a plain chat.
Tell the AI who you are. Example: "We are a state trade association for community bankers. We run an annual conference, chapter meetings, and continuing education. Our team is 9 people."
Paste or type any background the AI should know, program details, recent reports, grant guidelines. Write it in plain sentences, not code. Example: "Our spring conference drew 640 attendees last year. Our member renewal rate is 87%."
The AI will use this date when discussing timelines or deadlines.
Tell the AI how to talk to you. Example: "Explain things simply. Ask me questions if you need more information. Never invent facts."
Lets the model walk the site's tutorials, tools, topics, and their relationships when answering.
You can change any of these at any time. If you change them after chatting, earlier answers may need a second look.
All prior messages count toward the visible context budget. The conservative UTF-8 byte estimate is not exact tokenization. Over-budget conversations are blocked, never silently shortened.
Not sure what fits? Pick your graphics card to see which quantized models fit with headroom. Then load one in the browser, upload its GGUF, or install it in Ollama/LM Studio.
Run the setup sequence: plan the fit-check and harness install before downloading.
No connection attempted.
Your server must already be running and permit this exact origin. Browser local-network permission may also be required. Keep it on 127.0.0.1; do not enable wildcard origins or LAN access. The lab changes no server settings.
AssociationAI is AssociationAI's hosted assistant. Nothing is installed: the model runs on the hosting provider's computers, reached through AssociationAI's own connection. For each question, AssociationAI's server also looks up the most relevant passages from its published articles and tutorials and gives them to the model as reference material, so answers can point you to the piece they came from.
Your text leaves this device. Anything you type is sent to the hosting provider and may be logged by them. Use public or synthetic information only, never paste member, registrant, or sponsor records. For real work, use one of the on-device providers below; they never send your text anywhere.
Guests get a few free hosted messages a day with no sign-in at all. When the daily limit runs out, the hosted option pauses until the next day.
Every hosted chat is kept on record. Hosted chat is guest-only, with no sign-in, up to a small daily limit. Chats are recorded under a hashed identifier, not your name. All hosted chat records are retained for 12 months, and AssociationAI will provide them to authorities if a chat breaks the law. Never type anything into the hosted chat that you would not want kept on record.
Local runs WebLLM 0.2.85 in your browser, six curated mlc-ai models. Load downloads weights from external hosts, then browser inference runs through WebGPU. No model account is required. Unload frees the browser's GPU memory; switching models or providers only disconnects and keeps the current one warm.
Upload GGUF runs any GGUF file you provide via wllama (llama.cpp in the browser, WebGPU-accelerated). The file never leaves your device. Browser tabs cap out around 4 GB, so this fits models up to ~8B quantized; larger models belong on Ollama or LM Studio.
Qwen3.8 27B can use your installed Ollama or LM Studio model. Unload frees the model's GPU memory on your server (Ollama) or unloads its instance (LM Studio). No automatic model installation, weight download, API-key storage or remote fallback.
Up to 12 conversations, including sources and text, are stored in this browser profile. Anyone using the profile can read them. No cloud sync. Turning this off removes stored conversations but keeps this tab's working copies.
Previous workspace versions stay untouched until you explicitly import one.
Automatic spot-checks, not a fact-check. They catch mechanical mistakes, a number copied next to the wrong word, a missing citation, but they cannot tell you whether a claim is true. Read every claim yourself before you use it.
The lab is rung one. The kit takes you to rung two: a step-by-step guide to installing Ollama and running larger models (8B and up) on your own machine, plus a prompt pack written for association work, session descriptions, renewal nudges, board summaries, tuned for smaller local models.
Want the files behind this lab, sample dataset included? Email ai@trihelix.co and we will send them your way.