LAUNCH
Kyros Labs · 3 min read
Kepler 1.2 is live — our most powerful model yet
1.1 taught coding agents to gate tool calls. 1.2 expands to full System One: triage, moderation, incident noul, and the gates you already know.
Now available
Kepler 1.2Our most powerful model yet.
Kepler 1.2 is Kyros Labs’ open, local System One decision model — choice, score, and noul in one pass. Free to run on your machine. Fine-tuned from open Laya. Built for software that needs yes/no, route, and score — not another chatbot.
Model lineup
Shipped
Kepler 1.1
Coding-agent tool gates. Allow / ask / deny before shell, files, and network. Secret tripwire.
Previous weightsNow available
Kepler 1.2
Our most powerful model yet. Full System One — triage, moderation, incident noul, plus the gates from 1.1.
Weights on Hugging FaceComing soon
Kepler 1.3
An order of magnitude beyond any previous Kepler. Early access waitlist open.
Join waitlistSupport triageTool-call gatesTicket routingUrgency noulAgent guardrailsModerationIncident detectSecret tripwire
Support triageTool-call gatesTicket routingUrgency noulAgent guardrailsModerationIncident detectSecret tripwire
Support triageTool-call gatesTicket routingUrgency noulAgent guardrailsModerationIncident detectSecret tripwire
Support triageTool-call gatesTicket routingUrgency noulAgent guardrailsModerationIncident detectSecret tripwire
Reflex latency you can budget for
One forward pass. No paid cloud round-trip. Numbers that belong in a product SLA.
0ms
Typical local latency
One forward pass. No round-trip to a paid cloud API.
0$
Self-hosted cost
Open weights. Run on your GPU or CPU. Forever free.
0+
Languages via Laya stack
Multilingual routing when you need global coverage.
Most AI writes paragraphs.Kepler makes choices.
Pass a state and typed questions. Get choices, scores, and calibrated probabilities in one pass — ready for production control flow.
Reflex-speed
Milliseconds on commodity hardware.
No hallucination prose
Outputs are typed. nothing to regex.
Open & inspectable
Apache lineage. Weights you can host.
Air-gap ready
Keep sensitive state on-prem.
kepler · 1.2 · system one
State
{
"tool": "shell",
"command": "rm -rf /",
"cwd": "/workspace",
"agent": "cursor",
"goal": "cleanup"
}Question
choice → allow / ask / deny
Answer
1 passactiondenyconf 0.75
Blocks wipe commands before they run.
From the live Kaggle sanity check on open weights. Runs free on your machine.
Open reflex speed. No vendor meter.
Kepler is the open System One path: local, inspectable, and priced at zero for self-host. Not affiliated with TypeSafe AI or Jev.
| Metric | Kepler 1.2 | Jev (proprietary) |
|---|---|---|
| Typical local latency | ~33 ms | Cloud round-trip |
| Self-host cost | $0 weights | Vendor priced |
| Output shape | Typed choices | Vendor API |
| Inspectability | Open weights | Closed stack |
| Air-gap | First-class | Depends on vendor |
Everything you need to evaluate Kepler.
Coming soon
Kepler 1.3An order of magnitude beyond anything we’ve shipped.
If Kepler 1.2 goes viral — or you just want first access — join the waitlist. Name, email, country. That’s it.
Notes from the lab
Launch notes, System One design, and how Kepler fits real product code.
PRODUCT
Kyros Labs · 4 min read
Why typed decisions beat chat for control flow
Kepler returns choices and scores in one pass, so your product code never has to parse a paragraph.
RESEARCH
Engineering · 6 min read
Apache lineage, local weights, no vendor lock
Kepler is fine-tuned from open Laya. Host it yourself, inspect it, and keep sensitive state on-prem.
Questions teams ask before they ship
Short answers. Full detail lives in docs and the open repo.
Read docsIs Kepler really free?
Yes. Self-host the open weights at $0 from Hugging Face. You pay for your own compute. not a usage meter from Kyros Labs.
Where do I download the model?
Public weights: huggingface.co/MAKALY/kepler-1.2. pip install laya, then load that model id. Kepler 1.1 remains available for agent-gate specialists.
What changed from 1.1 to 1.2?
Kepler 1.1 focused on coding-agent tool gates (allow / ask / deny) and secret tripwires. Kepler 1.2 is our most powerful release — same gates, plus support triage, moderation, and incident noul across the full System One surface.
How is this different from a chat model?
Chat models write prose. Kepler answers typed questions. route, score, yes/no. so your code can branch without regex gymnastics.
How does Kepler compare to Jev?
Jev is a proprietary System One stack. Kepler is open, Apache-lineage via Laya, and designed to run locally with reflex-speed latency on commodity hardware.
What’s coming in Kepler 1.3?
Broader training and harder evals — built to be an order of magnitude beyond any prior Kepler cut. Join the waitlist for early access.
Ship the open alternative.
Open the playground. Download free weights. Run Kepler 1.2 where your data already lives.
Built on Laya (Apache 2.0). Not affiliated with TypeSafe AI or Jev.