Commercial
Everything we ship is free open source.
multivon-eval and pdfhellare both Apache 2.0 on PyPI today. No paid tier, no usage limits, no telemetry, no signup. That’s the whole pricing page.
pip install multivon-eval44 evaluators across deterministic, LLM-judge (QAG), agent-trace, compliance, multimodal, conversation, and consistency categories. Calibrated thresholds, hash-chained audit logs, pytest plugin, CLI.
Read more →pip install pdfhell30 adversarial PDFs across 3 trap families that break AI document readers. Procedural ground truth — no LLM-as-judge. JUnit + audit-pack output for CI.
Read more →What about custom / on-prem / enterprise?
Multivon is pre-product and pre-revenue. We’re not selling SaaS, not running a hosted service, and not packaging an “Enterprise” tier. The OSS is where the work lives today.
If you need something we don’t ship — adversarial PDFs procedurally generated from yourdocument templates, a custom trap family for your vertical, on-prem deployment guidance, audit-pack consulting, paid integration support — we’d love to hear what you need. We’ll figure out whether we can help and what it costs together.
Honest version: a few inbound conversations would tell us what people actually want to pay for, and we’d build the right thing rather than guessing.
Things you can already do with the OSS — for free, today
- ●Run
pdfhell run --model anthropic:claude-sonnet-4-6 --suite miniagainst any vision model and get a real pass-rate breakdown across 3 trap families. ~$0.01 per run. - ●Generate a tamper-evident audit pack (hash-chained ZIP with SHA-256 manifest) via
--audit-pack out.zip. Procurement teams accept this directly. - ●Gate your CI on pass-rate with
--fail-threshold 0.85and JUnit XML output via--junit results.xml. - ●Wire
multivon-evalinto pytest, evaluate RAG faithfulness, grade agent trajectories, run on-prem judges via OpenAI-compatiblebase_url. - ●Read every benchmark number on the leaderboard with a reproducer command and a raw-JSON source.