EXAMPLE WORKLOADS(14 CANONICAL MISSIONS)
AUTOCYCLE [01/14]
Compare engines
Find the best serving engine forQwen3-32B on vLLM / H100
Serve reasoning
DeployDeepSeek R1 Distill 8B for reasoning
Decode reality
Run
Physical Intelligence pi0 on Edge Mesh
Physical Intelligence pi0 on Edge MeshOptimize workload
OptimizeLlama 3.1 8B for <40ms TTFT latency
Frontier AI Lab & Reality Infrastructure.
HOROS builds frontier intelligence systems that learn from real-world evidence—discovering gaps, adapting models, running experiments, verifying outcomes, and evolving continuously.
Engineer intelligence against reality.
HOROS connects models, data, context, runtime, compute, workloads, and real-world evidence into one continuous system for building, testing, verifying, deploying, and improving intelligence.
WORLD STATE
Decoded, not guessed
SERVING STACK
Discovered per mission
WORKLOAD BENCHMARK
Tested on real traffic
GOVERNANCE CONTRACT
Bounded by authority
DEPLOYMENT KIT
Run it or export it
EVIDENCE BUNDLE
Verifiably proven
INDEPENDENT VERIFIER
Verified, not assumed
COMPOUNDED LESSONS
Learned from reality
ROOT CAUSE SRE
Traced to the hardware
FACILITY POWER
Optimized for megawatts
WORLD DYNAMICS
Fused across sensors
CODE REPAIR
Patched and verified
+
+
Reality Intelligence Runtime
From intent to a
verified reality you own.
Teach intelligent systems to decode reality, explore unknowns, autonomously experiment across models and hardware, and deliver verified production results.
+
+
Tell HOROS what you want
DeepSeek V3Llama 3.1 70BQwen 2.5 7BClaude 3.5 SonnetPhi-4Mistral Large 2Whisper Large v3
Gemma 2 9B
to do in plain English.
DeepSeek V3
Llama 3.1 70B
Qwen 2.5 7B
Claude 3.5 Sonnet
Phi-4
Mistral Large 2
Whisper Large v3
Gemma 2 9BHOROS explores
vLLM-OmnivLLM
SGLangTensorRT-LLMDiffusers
TEINVIDIA Dynamo
and instantly builds the
perfect system for you.
vLLM-Omni
vLLM
SGLangTensorRT-LLM
Diffusers
TEINVIDIA Dynamo
It tests and
benchmarks
everything in the background,
so you don't have to.
Run it anywhere, from
NVIDIA L40Sto
your own hardware,
and own the final result.
Decode
Sensor Telemetry
into verified, live
World State.
Acts within authorized
safety boundaries,
never exceeding your
organization's policy.
Proves every outcome
through independent
verification and
empirical evidence.
Turns verified results
into compounding
intelligence that learns
from every mission.
New Session
Optimize Llama-3.1-8B on vLLM for latency
>Thought for 1s
Capturing cost-first intent for Llama 3.1 8B on vLLM.
Intake updated
Model, engine, goal
ModelLlama 3.1 8B
EnginevLLM
GoalLow cost, latency checked
Plan drafted3.1s
>
10 execution phases prepared
Plan ready
review / 10 phases / ~23mOptimize Llama 3.1 8B on vLLM for latency
Recommended path: vLLM on L4. Review plan before execution.
Working...
Mission Runtime Execution Plan
HOROS Reality Intelligence Loop
draft
Primary Intent
Optimize Inference Cost
Hard Constraints
P99 <= 250ms, 24GB
Verification Mode
Empirical Benchmark
Execution & Evolution Plan
15 phases, 3 validated
01
Input Normalizationready
Process text, telemetry & sensors
02
Outcome Intelligenceready
Extract intent & build contract
03
Capability Discoveryready
Find models, runtimes, hardware
04
System Composerqueued
Build candidate stacks
05
Experiment Enginequeued
Simulate, sandbox, benchmark
06
Autonomous Decisionqueued
Quality and economics ranking
07
Policy & Authorityqueued
Risk, cost, and safety gates
08
Execution Layerready
Deploy software or actuate physical
09
Evidence Collectionqueued
Observe and gather provenance
10
Independent Verifierqueued
Verifiably confirm claims
11
Experience Enginequeued
Compile verified outcome records
12
Lesson Enginequeued
Generalize into teaching signals
13
Teaching Systemqueued
Train, adapt, and route models
14
Intelligence Evolutionqueued
Evaluate, canary, and promote
15
Better Reality Modelqueued
Continuously adapt world state
Review the mission contract, then authorize.
HOROS CANONICAL_V4
OPTIMIZATION RESULTS
completed on L40S
Applied Optimizations
3 phases verifiedModel
AWQ int4 quantizeSpecialize
Runtime
FlashAttention v2Paged KV cacheCUDA graphs
Compute
GPU Auto-TuneL40S placement
Latency
-79%38.2 ms
baseline 184ms
measured across 10k reqs on L40S
TTFT p50
-68%24.1 ms
baseline 75ms
prefix caching + AWQ kernel prefill
TPOT p50
+185%4.2 ms/tok
baseline 12ms
FlashAttention v2 + CUDA Graphs
Throughput
+216%518 tok/s
baseline 164 tok/s
continuous batching @ 64
Cost per 1M tokens
-71%$0.12
baseline $0.41
L40S ($0.92/hr) vs A100 baseline
VRAM
-57%14.2 GB
baseline 33.1 GB
fits in single 48GB GPU budget
Confirmed winner: NVIDIA L40S
94.2s runtime elapsed
Deploy, Apply, or Actuate
ready
Select target across Computational, Edge, Physical, or Industrial systems.
TARGETHARDWARE / GPUS
HOROS Cloud
Fully managed endpoint, auto-scaling
H100L40SA100L4
Your RunPod
Deploy to your own RunPod account
H100A100RTX 4090L40S
Your Modal
Deploy to your own Modal workspace
H100A100L40ST4
Kubernetes (K8s)
Native Helm charts & custom operators
L40S PoolA100 NVLink
GENERATED ARTIFACT
1FROM horos/vllm:0.6.3-l40s
2ENV MODEL=Llama-3.1-8B-Instruct
3ENV QUANTIZATION=awq_marlin
4COPY ./weights /models
5CMD vllm serve $MODEL \
6--quantization awq_marlin \
$0.12 / 1M38ms p95
Decode reality· VEHICLE-07
live stream98 FPS
Ingesting Camera-01, Camera-02, LiDAR-01, and CAN-Bus telemetry.
SPATIAL RADAR & DETECTED ENTITIESDEPTH: 98.4% CONF
pedestrian
14.2m98% conf
cyclistparallel northbound
21.1m94% conf
vehiclerel speed -8 km/h
31.4m99% conf
Speed: 42 km/hRoad: wet surfaceVisibility: moderate
PREDICTED NEXT STATE:pedestrian likely enters path in 1.4s
Camera-02 partially occluded (Uncertainty: 0.38)
Fallback to LiDAR-01Deliverable: WorldState:v4
Policy & Safety Authority
Autonomous governance, blast radius checks, and spend boundary enforcement
passed
AUTONOMY LEVEL
L3 Bounded Prod
SPEND AUTHORITY
$340 / $500 cap
SAFETY GATES
6 / 6 Validated
SAFETY GATES (A–F)
6 autonomous checks validated
01
Gate A · Objective Ambiguitypassed
Explicit intent: Confidence score 0.96
02
Gate B · Data Residencypassed
EU-West-1 region enforcement validated
03
Gate C · Spend Authoritypassed
Projected $340/mo within $500/mo hard cap
04
Gate D · Blast Radiusisolated
Isolated to staging-cluster-02 namespace
05
Gate E · Destruction & Reversalverified
Non-destructive plan with rollback snapshot
06
Gate F · Reality Freshnessclean
Zero conflicting world state facts found
Rollback Snapshot: WorldState:v3.1.8-clean
Pre-authorized execution
Independent Proof & Verification
The executor is never the sole judge of its own success
verified
VERIFIER VERDICT
100% PROVEN
OBSERVATION WINDOW
Active (94.2s)
CROSS-VALIDATION
4 / 4 Sources
INDEPENDENT SOURCES
Cross-validated telemetry
01
AWS Cost Explorer (CUR)
verified-$34.20/day verified delta match
02
OpenTelemetry P95 Collector
verified38.2ms p95 across 10,000 trace samples
03
HOROS Golden Eval Suite
verified98.9% accuracy with zero quality regression
04
NVIDIA NVML Power Sensors
verified185W average sustained on L40S
05
Audit Provenance Trace
signedVerifiable SHA256 immutable receipt
06
Independent Verifier Signature
validHOROS-Independent-Verifier-04
sha256:7f8a9e21d3c544e98f0291ba4c90184b2c198a0029b47e
Teaching Fabric & Intelligence Evolution
EXPERIENCELESSON ENGINESYNTHETIC DATA
compounding
STATE UNDERSTANDING
64% → 88% (+24%)
ANOMALY ACCURACY
51% → 79% (+28%)
COST EFFICIENCY
$0.12/1M (-71%)
SYNTHETIC DATA FACTORY / LESSONS
Software Cloud Physical
VERIFIED EXPERIENCE
ACTIVE LEARNING
SYNTHETIC DATA
01
MoE Burst Routingpromoted
Prefer AWQ-Marlin on L40S for 79% latency drop
02
FP8 KV Page Allocationshadow
Eliminates memory fragmentation during peak
03
State Perception v4promoted
Understanding accuracy improved 64% → 88%
04
Anomaly Detection Ruleevaluate
Detection accuracy compounded 51% → 79%
05
Autonomous Policy Tuningpromoted
14 matching production workloads updated
06
Background Drift Checkshadow
Zero regression across golden eval suite
Drift Monitor: 0 material drift events
EVALUATESHADOWCANARYPROMOTECONTINUOUS AGENT
+
+
FRONTIER AI LAB · REALITY INTELLIGENCE INFRASTRUCTURE
Teach intelligence to decode your reality
TEACH CONTINUOUSLY
Adapt models to your physics
Provide HOROS with open models and physical telemetry. The runtime continually adapts weights, state, and rules to match your environment.
INTELLIGENCE ROUTING
Route intent, not inference
Stop manually calling model APIs. Send intent to the runtime, and it routes tasks to the best adapted intelligence for the job.
Connect with any sensory hardware
Size the GPU / XPU edge target
Allocate shared cache & VRAM limits
Qwen 2.5 Coder 32BAVAILABLE
32B DenseCode IntelAST Verifiers<35ms TTFT
Qwen/Qwen2.5-Coder-32B-Instruct
Specialized code generation, repository refactoring, automated testing & contract synthesis
Served via 
Modal, AWS & Edge Mesh
$0.08
per 1M cached tokens
$0.24
per 1M output tokens
131,072 tokens
Context window
Built and served with
NVIDIA
vLLM
PyTorchTensorRT-LLM
Docker
Kubernetes+
+
+
+
Deploy and verify anywhere
HOROS deploys intelligence into software, compute clusters, edge devices, and physical robotics. Actuate when authorized, and verify the physical consequences.
Hugging FaceAny Hugging Face modelHOROS
HOROS Cloud
RunPod
V
Vast.aiLocal hardware
+
+
Our cloud, your cloud, or the physical edge
The Constant
Intelligence runs everywhere identically. You own the verified outcomes.
Managed
HOROS Cloud
A fully managed Reality Intelligence Runtime. HOROS handles node provisioning, container orchestration, and continuous state synchronization. You simply provide the authorized intent, and we scale the capabilities securely, billed purely on verified usage.
Your Cloud
RunPod, Modal, Vast & CoreWeave
Bring your account, your endpoint URL, and your key. The Execution Layer seamlessly bridges to your preferred GPU cloud.
We manage the complexity of deploying models and tools into these environments, orchestrating the nodes and scaling the stack securely within your own infrastructure boundary.
More clouds coming soon
Your Hardware
Local Hardware & Robotics
The deployment kit runs the exact same intelligence stack on anything you own. We manage the synchronization and state whether you deploy to AWS clusters, Azure confidential compute, NVIDIA edge GPUs, AMD APUs, or physical robots.
Read the security page
+
+
Own the winning stack before you ship
Not a black box. You get the measured stack to run, deploy, or export.
Supported across the stack
Open models, serving engines, GPUs, and the clouds you deploy to. HOROS supports every layer of the inference stack.
MODELS
IMAGEQwen-Image
EMBEDNV-Embed
VIDEOCosmos
VISIONPixtral
LLMDeepSeek-V3
LLMLlama-3.1
AUDIOWhisper
PHYSICAL1X NEO
WORLDFSD v12
IMAGEQwen-Image
EMBEDNV-Embed
VIDEOCosmos
VISIONPixtral
LLMDeepSeek-V3
LLMLlama-3.1
AUDIOWhisper
PHYSICAL1X NEO
WORLDFSD v12
IMAGEQwen-Image
EMBEDNV-Embed
VIDEOCosmos
VISIONPixtral
LLMDeepSeek-V3
LLMLlama-3.1
AUDIOWhisper
PHYSICAL1X NEO
WORLDFSD v12
ENGINES
ENGINEvLLM
ENGINETensorRT-LLM
ENGINEvLLM Omni
ENGINEvLLM
ENGINETensorRT-LLM
ENGINEvLLM Omni
ENGINEvLLM
ENGINETensorRT-LLM
ENGINEvLLM Omni
GPUS
48 GBL40S
24 GBRTX 4090
80 GBA100
80 GBH100
141 GBH200
192 GBB200
24 GBL4
16 GBT4
48 GBL40S
24 GBRTX 4090
80 GBA100
80 GBH100
141 GBH200
192 GBB200
24 GBL4
16 GBT4
48 GBL40S
24 GBRTX 4090
80 GBA100
80 GBH100
141 GBH200
192 GBB200
24 GBL4
16 GBT4
CLOUDS
A100RunPod
B200HOROS Cloud
L40SAWS
H200Azure
A100RunPod
B200HOROS Cloud
L40SAWS
H200Azure
A100RunPod
B200HOROS Cloud
L40SAWS
H200Azure
DATACENTER
FACILITYData Center
COMPUTEBare Metal
COMPUTECPU/XPU
INFRAMemory
INFRANetwork & Routing
INFRAStorage
SCALEKubernetes (K8s)
SCALEScaling
FACILITYData Center
COMPUTEBare Metal
COMPUTECPU/XPU
INFRAMemory
INFRANetwork & Routing
INFRAStorage
SCALEKubernetes (K8s)
SCALEScaling
FACILITYData Center
COMPUTEBare Metal
COMPUTECPU/XPU
INFRAMemory
INFRANetwork & Routing
INFRAStorage
SCALEKubernetes (K8s)
SCALEScaling
EDGE
PHYSICALRobots & Vehicles
PHYSICALMachines
EDGELocal Compute
EDGEGateways
EDGECameras
EDGESensors
OBSERVETelemetry
OBSERVEStreams
PHYSICALRobots & Vehicles
PHYSICALMachines
EDGELocal Compute
EDGEGateways
EDGECameras
EDGESensors
OBSERVETelemetry
OBSERVEStreams
PHYSICALRobots & Vehicles
PHYSICALMachines
EDGELocal Compute
EDGEGateways
EDGECameras
EDGESensors
OBSERVETelemetry
OBSERVEStreams
+
+
Common questions
Can't find what you're looking for? Get in touch
What is HOROS?
HOROS is a Frontier Intelligence Engineering System. It connects models, data, context, runtime, compute, workloads, and real-world evidence into one continuous loop for building, testing, verifying, deploying, and evolving intelligence against reality.