CarraZero Models
Nine specialized models across four families — rerankers, embeddings, custom fine-tunes and the runtime that serves them. Specs, latency and pricing, in the open.
Reranker
3 models
Embedding
3 models
Custom
2 models
Inference
1 model
Catalog
Pick the smallest model that clears your quality bar. We will tell you which one that is after the benchmark — including when the answer is a model you already run.
CarraRank-2
carrarank-2
Flagship cross-encoder reranker. Highest accuracy on messy enterprise corpora — contracts, tickets, policies and long-form documents.
Best for: Enterprise search, RAG, agent retrieval
CarraRank-1
carrarank-1
Balanced reranker for high-QPS workloads. Near-flagship quality at roughly half the compute per query.
Best for: High-volume search, voice + real-time agents
CarraRank-Mini
carrarank-mini
Edge-class reranker for latency-critical paths. Runs comfortably on CPU or a shared GPU slice.
Best for: On-device, in-VPC, ultra low-latency
CarraEmbed-L
carraembed-l
Domain-tuned dense embeddings for enterprise corpora. Trained to separate near-duplicate policy and contract language.
Best for: Primary index for RAG and semantic search
CarraEmbed-S
carraembed-s
Compact embeddings for large indexes where storage and recall cost dominate. Drop-in with the same tokenizer as CarraEmbed-L.
Best for: Billion-scale indexes, first-stage recall
CarraEmbed-Code
carraembed-code
Code and runbook embeddings trained on repositories, SOPs and internal wikis — symbol-aware, not just prose-aware.
Best for: Internal code search, engineering agents
CarraTune-Extract
carratune-extract
Structured extraction fine-tunes: pull typed fields out of claims, invoices, leases and statements with schema-constrained decoding.
Best for: Claims, underwriting, AP/AR, lease abstraction
CarraTune-Classify
carratune-classify
Routing, triage and scoring models trained on your own labels and outcomes — deterministic, cheap and auditable.
Best for: Ticket routing, intent, risk and lead scoring
CarraServe
carraserve
The serving runtime behind every model above: batching, quantization, KV reuse and autoscaling tuned for enterprise concurrency.
Best for: Anywhere you need the weights close to the data
Lifecycle
01
Benchmark
We run your corpus and queries against your current stack and every CarraZero candidate.
02
Tune
Hard-negative mining and training against your judgements, not a public leaderboard.
03
Evaluate
A frozen golden set, versioned with the model, so quality regressions are visible before release.
04
Deploy
One API, your cloud, your VPC or on-prem — with shadow traffic before cutover.
Pricing
Rerankers
from $0.04
per 1k documents ranked
Volume tiers from 10M documents/month. Mini is priced at a third of flagship.
Embeddings
from $0.012
per 1M tokens
Batch indexing discounts apply above 500M tokens/month.
Custom models
from $12k
per model, then usage
Includes data curation, training, eval harness and two tuning rounds.
Self-hosted
Annual licence
unlimited inference
Open weights in your VPC or on-prem. You keep the weights after term.
Indicative list pricing. Committed-use and self-hosted terms are quoted after the benchmark, once we know which model you actually need.
API
Rerank
from carrazero import CarraZero
client = CarraZero()
ranked = client.models.rerank(
model="carrarank-2",
query="Which vendor contracts renew in Q3?",
documents=docs,
top_n=10,
)Embed
from carrazero import CarraZero
client = CarraZero()
vectors = client.models.embed(
model="carraembed-l",
inputs=chunks,
dimensions=768,
input_type="document",
)We'll pick the model, run the benchmark on your corpus and show you the numbers in a week.
Schedule your intro call→