Stop paying LLM prices for every decision.
Sieve turns context into clear choices with confidence scores. Decisions come back faster and cheaper, with no generated text to parse.
Decision Index score for Sieve-31B-v3
On the Decision Index, October 2026
Median time per decision for Sieve 31B v3
The 31B model fits a single 96 GB card
One pass gives one probability per answer.
Complete claim documents. Low claimed amount. No prior fraud indicators.
Open decision models, 0.8B to 31B parameters
Pick the one that fits the job. Smaller models cost less to run and are usually quicker per decision at the low end, while larger ones handle harder reasoning. Every model returns its answer in the same format, so you can switch between them later without rewriting anything.
| Model | Best at | Decision Index (v0.4.1) | Median latency | Runs on |
|---|---|---|---|---|
| Sieve 31B v3 | Hard reasoning and agent tool selection | 71.1 · #1, tied | 232 ms | One 96 GB GPU |
| Sieve 27B v2 | Intent detection, claim verification, and RAG hallucination checks | 59.9 | 106 ms | One 80 GB GPU |
| Sieve 9B v2 | The balance of accuracy, speed, and cost | 47.0 | 51 ms | About 20 GB of GPU memory |
| Sieve 9B | The previous 9B release, kept for teams already running it | 41.7 | 50 ms | About 20 GB of GPU memory |
| Sieve 4B | Best-calibrated confidence scores: 3.1-point calibration error, the lowest in the family | 41.0 | 93 ms | About 10 GB of GPU memory |
| Sieve 2B v3 | Lightweight intent detection and routing | 34.5 | 43 ms | About 5 GB of GPU memory |
| Sieve 2B | The previous 2B release | 21.9 | 63 ms | About 5 GB of GPU memory |
| Sieve 0.8B | The smallest footprint, for edge and high-volume routing | 20.8 | 52 ms | About 2 GB of GPU memory |
Decision Index scores are from edition 0.4.1, board data of 28 October 2026. The board takes new entries continuously, so ranks and totals move. Median latency is the time taken per decision; the Sieve 31B v3 figure is the Decision Index measurement on one NVIDIA RTX PRO 6000.
Your industry has its own knowledge. Train Sieve on it.
Decision Sets are industry-specific training data: the examples, terminology and correct answers your agents need.
A pharmaceutical company's agents work with different knowledge than a bank's. Upload your labeled examples in Foundry and customize Sieve for your domain.
Upload your Decision Set
Bring labeled examples from your industry.
{"state": "Adalimumab 40mg, plaque psoriasis,
two prior biologics failed, step
therapy documented",
"question": "prior_auth",
"options": ["approve", "deny", "request_records"],
"label": "approve"}
Customize a Sieve model
Choose an approved base model and training settings.
Sieve Flash · Supervised LoRA
Test it, then put it to work
Evaluate in Playground. Call your model through the API.
Build agents that decide better.
Start with Sieve. Make it yours when you're ready.