Open-Weight 1.6B V-JEPA Model Access Inquiry
Safetensors weights for our 1.6B Spatial AI World Model are available under the Apache 2.0 license. Submit a link inquiry below for weight access and technical documentation.
KS-WI Open Model & Production Engine
KineSpatial World Intelligence (KS-WI) is a 1.6B parameter V-JEPA latent world model. Access open weights via direct link inquiry or license the optional Hardened Production Engine.
Open-Weight Community vs. Production Hardened Enterprise Tier
We believe core spatial intelligence should be accessible to all AI researchers. The open-weight model is distributed freely on Hugging Face. For physical hardware deployments requiring low latency and hard safety guarantees, our commercial Production License provides hardened compiled runtimes.
Open-Weight Hugging Face Distribution
- Full 1.6B V-JEPA Dense Safetensors Checkpoint
- PyTorch & Apple MLX Inference Scripts
- Hugging Face Hub `transformers` Pipeline Support
- Open Research & Non-Gated Prototyping
Hardened & Optimized Runtime Variant
- Bare-Metal C++ PREEMPT_RT (`SCHED_FIFO 99`) Runtime
- TensorRT FP8/INT8 Quantized Engine (< 1.2ms latency)
- ASIL-D Compliant Hardware Safety Torque Clamp
- Dedicated Systems Engineering SLA & Fine-Tuning Support
Architecture & Overview
KS-WI provides physical robotics with a predictive imagination engine. Available under the Apache 2.0 License via link inquiry, researchers and developers can run latent contact predictions directly in PyTorch or MLX. (TODO: Direct Hugging Face Hub automated weight repository association). For enterprise hardware deployments requiring real-time sub-millisecond guarantees, we offer a commercial Hardened PRODUCTION-equivalent License featuring pre-compiled C++ PREEMPT_RT zero-copy binaries, TensorRT FP8 quantization, and hardware torque boundary clamps.
Key Technical Deliverables
- Open-Weight safetensors model checkpoints available via link inquiry.
- Seamless integration via PyTorch / MLX latent prediction loops.
- Hardened PRODUCTION Tier available with zero-copy C++ PREEMPT_RT runtime binaries.
- TensorRT FP8/INT8 quantized engine graphs for sub-millisecond execution.
- ASIL-D compliant hardware torque clamps to prevent physical boundary breaches.
# Open-Source Model Loading (Apache 2.0)
# TODO: Direct Hugging Face Hub automated repo association (luminus-ai/ks-wi-1.6b-jepa)
from huggingface_hub import hf_hub_download
import mlx.core as mx
from ks_wi import SpatialWorldModel
# Fetch 1.6B open weights (Link inquiry access token until automated HF association)
# weights_file = hf_hub_download(repo_id="luminus-ai/ks-wi-1.6b-jepa", filename="model.safetensors")
model = SpatialWorldModel.from_pretrained("luminus-ai/ks-wi-1.6b-jepa")
# Run latent spatial prediction horizon
latent_state = model.encode_perception(camera_feed, imu_vector)
predicted_horizon = model.predict_horizon(latent_state, steps=10)
print(f"[+] KS-WI Open Model prediction complete: {predicted_horizon.shape}")
# Enterprise Production Upgrade Option:
# engine = model.compile_hardened_runtime(tier="PRODUCTION", scheduler="PREEMPT_RT_99")