jean-rec-1

Closed beta. Reach out for a key.

A pretrained generative recommender you can call without building a model first. You fit a tokenizer on your catalog, send interaction histories, and get recommendations back. There is no architecture to choose, no embedding table to size, and no six-week ramp before the first useful result.

Retrieval and ranking are collapsed into one model that emits item identifiers directly, rather than a candidate generator bolted to a reranker. One system to operate, one objective to tune, and no stage boundary for quality to leak across.

Out of the box

The useful property of a foundation model is that the day-one result is not zero.

	Classical recommender	jean-rec-1
Before first result	Collect interactions, train, tune	Fit a tokenizer, send a history
New item	Unusable until it has interactions	Recommendable as soon as it has content
New tenant	Starts from scratch	Starts from the pretrained backbone
Retrieval and ranking	Two systems	One model

You still get better results after adapting it on your sequences. The point is that you do not have to wait for that to see whether it works.

Specification

Task	Next-item prediction over a catalog
Architecture	Decoder-only transformer over semantic ID tokens
Item vocabulary	4-level multi-resolution codebook, see Semantic IDs
Context	512 interactions, truncated from the front
Input	An ordered interaction history
Output	Semantic ID sequences, decoded under a constrained beam to real items
Adaptation	Supervised fine-tuning on your sequences, then RL against your ranking objective
Status	Closed beta

Calling it

Recommend from a catalog

curl https://api.jeantechnologies.com/v1/generate \
  -H "Authorization: Bearer $JEAN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "jean-rec-1",
    "tokenizer_id": "tok_9k2m",
    "history": ["sku_310", "sku_884", "sku_771"],
    "limit": 10
  }'

Next action, no catalog

curl https://api.jeantechnologies.com/v1/models/jean-rec-1/predict \
  -H "Authorization: Bearer $JEAN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "history": [
      {"action": "view",     "item_id": "sku_310", "ts": 1735689600},
      {"action": "add_cart", "item_id": "sku_884", "ts": 1736294400}
    ],
    "top_k": 5
  }'

Represent a person

curl https://api.jeantechnologies.com/v1/models/jean-rec-1/embed \
  -H "Authorization: Bearer $JEAN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "history": ["sku_310", "sku_884", "sku_771"]
  }'

See Generate, Predict, and Embed for the full request and response shapes.

What it is not good at

Worth saying plainly, because a model that claims everything is not a model anyone can plan around.

• Catalogs without content. The whole approach assumes items carry real text or images. Bare SKU strings degrade it badly.
• Very short histories. With one or two interactions it falls back on catalog priors and behaves close to a popularity ranker.
• Explaining itself. It emits an identifier, not a reason. If you need a human-readable rationale, that is a separate model on top.
• Anything about language. It has no opinion on your copy.

---

Request beta access: jonathan@jeantechnologies.com