Open systems for intelligence, agents, security, and AI infrastructure.
CONTINUUM · ORCAROUTER · ORCACYBER · PAPERS ↓ · HUGGING FACE · OLLAMA · X · DISCORD
RESEARCH · INFRASTRUCTURE · SECURITY · OPEN SOURCE
Intelligence is becoming infrastructure.
Continuum AI builds open systems for models, agents, security, observability, routing, and alignment.
We build at the layers between models and the real world:
MODEL → ROUTE → AGENT → TOOL → ACTION
↑ │
└────── EVALUATE ← VERIFY ← TRACE ─────┘
Our goal is simple:
Build intelligence that can be routed, observed, secured, reproduced, evaluated, and evolved.
Self-hosted infrastructure for multi-model AI.
OpenAI-compatible LLM routing with BYOK, streaming, model interoperability, and a managed safety net.
Time travel for AI agents.
Record an agent run. Replay it. Fork from any checkpoint. Change the model while keeping everything else constant.
Local and cloud traces speak the same format — giving agents a reproducible history across environments.
Change behavior. Don't change the weights.
Runtime behavioral intervention for compressed LLMs.
0 WEIGHTS MODIFIED
0 RE-QUANTIZATION
BIT-IDENTICAL ORIGINAL MODEL PACK
129 RESIDUAL INTERVENTION SITES
INFERENCE-TIME CONTROL
Built first for Ternary Bonsai 2 27B while preserving the original compressed model.
An open archive of how AI agents actually work.
A versioned, verifiable archive of real-world system prompts, developer instructions, tool schemas, and agent harnesses.
Built to make agent behavior easier to study, compare, reproduce, and understand.
AI code review with an enforcement layer.
An open multi-model code review harness for automated reviews, security analysis, severity-ranked findings, and merge gates.
CODE
↓
MULTI-MODEL REVIEW
↓
P0 · P1 · P2 · P3
↓
MERGE GATE
A public record of when AI systems fail.
An open, structured archive of real-world AI incidents — outages, safety failures, model regressions, security events, policy failures, and other production issues.
Built to make AI reliability easier to study over time across models, providers, and infrastructure.
CONTINUUM AI
│
┌────────────────────┼────────────────────┐
│ │ │
▼ ▼ ▼
ORCAROUTER ORCACYBER OPEN MODELS
│ │ │
AI GATEWAY SECURITY RESEARCH
ROUTING DEFENSE WEIGHTS
OBSERVABILITY AGENTS QUANTS
EVALS RESEARCH RUNTIMES
GUARDRAILS CODE SECURITY POST-TRAINING
│ │ │
└────────────────────┼────────────────────┘
│
▼
OPEN SOURCE
│
┌──────────────┼──────────────┐
▼ ▼ ▼
BUILD VERIFY EVOLVE
OrcaRouter is Continuum AI's infrastructure layer for operating models and agents across providers.
┌─────────────────┐
│ YOUR AGENT │
└────────┬────────┘
│
▼
┌───────────────────┐
│ ORCAROUTER │
│ │
│ ROUTE · OBSERVE │
│ SECURE · EVAL │
└─────────┬─────────┘
│
┌────────────────┼────────────────┐
│ │ │
▼ ▼ ▼
MODEL A MODEL B MODEL N
ADAPTIVE ROUTING · OBSERVABILITY · AGENT FIREWALL · GUARDRAILS · EVALS · BYOK · ZDR
OrcaCyber is our cybersecurity research and systems initiative for AI-native security.
Our work spans:
VULNERABILITY RESEARCH
AUTONOMOUS SECURITY AGENTS
CODE SECURITY
RED-TEAM RESEARCH
DEFENSIVE AI
SECURITY MODELS
AGENT SECURITY
SOFTWARE + AI SYSTEMS
│
▼
┌─────────────┐
│ ORCACYBER │
└──────┬──────┘
│
┌────────────┼────────────┐
▼ ▼ ▼
DISCOVER ANALYZE DEFEND
│ │ │
└────────────┼────────────┘
▼
VERIFY
We publish open-weight models, post-training research, quantizations, runtimes, and experimental releases for developers and researchers.
Models · weights · quantizations · research releases
Run Orca models locally.
We study the systems, architectures, and failure modes that emerge as AI moves from models into the real world.
Our work spans routing, alignment, security, robustness, agents, retrieval, and AI infrastructure — with an emphasis on research that can be tested, reproduced, and turned into working systems.
Zhenghua Bao · Fengya Tian · Chris Zhang · Zhenjun Chen · Xile Ma · Yi Shi
A production-oriented framework for intelligent model routing, combining contextual bandits with hybrid offline-online learning to continuously adapt routing decisions from real-world feedback.
MODEL ROUTING · CONTEXTUAL BANDITS · ONLINE LEARNING · AI INFRASTRUCTURE
Yi Shi · Tanyu Chen · Kai Shen
An empirical investigation into whether a remarkably simple white-box intervention can disrupt safety alignment at frontier scale — and what changes when alignment techniques designed for dense models encounter large mixture-of-experts architectures.
AI SAFETY · ALIGNMENT · MODEL SECURITY · MoE · INTERPRETABILITY
Zhenghua Bao
Examines how errors propagate across speech recognition and retrieval-augmented generation pipelines, showing how increasingly capable multi-hop retrieval can amplify upstream entity errors rather than eliminate them.
EMNLP 2026 · Main Conference
RAG · ROBUSTNESS · RETRIEVAL · SPEECH · MULTI-HOP REASONING
|
Model routing |
Model security |
|
Safety alignment |
Post-training |
|
Tool use |
Tracing |
|
Retrieval-augmented generation |
Reproducible experiments |
Research → Open Source → Production
We don't treat research as a separate layer from engineering.
When possible, our work becomes code, models, datasets, evaluations, or production systems that others can inspect, reproduce, and build on.
OPEN > CLOSED
MEASURE > ASSUME
REPRODUCE > DEMO
VERIFY > TRUST
SYSTEMS > WRAPPERS
EVOLVE > FREEZE
We publish the systems, artifacts, experiments, models, and research behind our work whenever possible.
The future of intelligence will be built across models, agents, infrastructure, and security.
Come build it with us.