The Memory and Control Layer
Multi-model AI systems break at the seams. Memory gets tied to one provider, routing turns into glue code, and interoperability depends on brittle wrappers instead of shared structure.
82d provides the memory and control layer: cross-model memory, semantic routing, and interoperability across heterogeneous AI environments.
"The per-token pricing model is going to look ridiculous in hindsight. Imagine paying per-word to read a book or search your own notes. That's where we are with embeddings and generation right now." — Andrej Karpathy @karpathy Paraphrased from 2024–2025 commentary on embedding economics
The Interoperability Tax
Four costs that compound when your AI stack spans multiple models, vendors, and environments.
Storage bloat
1536D vectors = 6 KB each. At 10M documents, that's 61 GB of coordinates — most of it redundant dimensions you never query directly.
Routing dependency
Every query and handoff depends on model-specific APIs, wrappers, and rate limits. Your continuity and control surface live on someone else's stack.
ETL bottleneck
Want Wikipedia in your RAG? 41.5M passages. $5,000+ to embed. 63 GB to store. Months to build the ingest pipeline.
Continuity break
text-embedding-ada-002 → sunset. Re-embed everything. Pay again. Rebuild compatibility. Repeat forever.
Cross-Model Memory and Routing
The near-term wedge: persistent memory, semantic routing, and interoperability across model changes.
Continuity across model changes.
- Cross-model memory: embeddings from OpenAI, Cohere, mxbai, nomic, MiniLM, and more land in one shared 82D space
- Semantic routing: one representation layer for search, handoff, and mixed-model retrieval
- Interoperability: merge datasets embedded by different teams or providers into one usable system
- Continuity: when a provider sunsets a model, your 82D coordinates stay stable
- 18.7× smaller: 6,144 bytes → 328 bytes per vector, reducing RAM, storage, and bandwidth costs
Quick Start
Project embeddings from any model into the same memory and control layer in one API call.
from eightytwo import Client
client = Client(api_key="your-key-here")
# Works with ANY embedding model
# OpenAI 1536D, Cohere 1024D, nomic 768D, etc.
vectors_1536d = openai_client.embeddings.create(...).data
vectors_82d = client.project(vectors_1536d)
# → model auto-detected from dimension
# Or specify the model explicitly
vectors_1024d = mxbai_client.embed(texts)
vectors_82d = client.project(vectors_1024d, model="mxbai-embed-large")
# Both land in the SAME 82D consensus space
# → directly comparable, permanently yours
print(f"Size: {1536*4}B → {82*4}B per vector = 18.7x smaller")Sign up to get your API key and endpoint URL.
# Project vectors to 82D consensus space
curl -X POST https://api.82d.ai/project \
-H "Authorization: Bearer YOUR_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"vectors": [[0.01, -0.02, ...1536 dims...]],
"model": "openai-3-small"
}'
# Response:
{
"vectors": [[0.04, 0.10, ...82 floats]],
"count": 1,
"input_dim": 1536,
"output_dim": 82,
"processing_time_ms": 0.3
}
# List supported models
curl https://api.82d.ai/modelsPaste 1536-dimensional vectors (from OpenAI, Cohere, etc.) to project to 82D.
Sign in above to enable live demo
Simple Pricing
Free during early access. Usage is monitored per account.
You own the coordinates. Nothing stored on our side unless you ask.
Early Access — Free Tier
Every account gets free monitored access to the full projection API, pre-built Wikipedia search, and a trained W matrix. Paid credit packs will be available when we leave early access.
Every account includes a trained W matrix + Wikipedia search access. Paid tiers coming soon.
Built For
RAG at Scale
Drop 41.5M pre-projected Wikipedia passages into your pipeline. Add your own embeddings from any model. One unified 82D index.
Model Migration
Moving from OpenAI to Cohere? Project both to 82D. Zero re-embedding. Zero downtime. Your existing vectors just work.
Multi-Team Search
Engineering uses mxbai. Research uses nomic. Product uses OpenAI. 82D makes them all searchable in one index.
Cost Control
18.7× smaller vectors = 18.7× less RAM, storage, and bandwidth. At scale, that's the difference between renting GPUs and not.
Beyond Projection
Projection is the foundation. On top of it: platform migration, pre-built datasets, agent interoperability, multimodal search, and managed infrastructure — nine services, one coordinate system.
Explore Services →Your vectors. Your coordinates. Your call.
First 10 MB free. No credit card. See the math for yourself.
Start Free