Private AI inference and confidential GPU computing infrastructure
Open-source engine for faster local language model inference with bit-for-bit identical output
Stage
Not given
Not given
Founded
Not given
Not given
Based in
US
Not given
Pricing model
Paid
Open source
Pricing
Confidential Inference: DeepSeek V4.1 Flash $0.30 input and $1.20 output per 1M tokens; Kimi K3 $3.00 input and $15.00 output per 1M tokens; GLM 5.3 $1.40 input and $4.40 output per 1M tokens; Qwen3.8 27B $0.30 input and $1.80 output per 1M tokens. Confidential VMs: B300 at $7.50 per GPU-hour, B200
Not given
Key features
OpenAI-compatible endpoint for private LLM inference
Hardware-encrypted VMs on Intel TDX, AMD SEV-SNP and NVIDIA GPUs
Client attestation verification before sending requests
Support for multiple open-weight models including DeepSeek, Kimi, GLM, Qwen
Confidential Kubernetes for data center scale deployments
Confidential training and fine-tuning on licensed stack
Root access over SSH to confidential VMs
Encrypted GPU communications with NVLink encryption on B200 and B300
Confidential agents runtime support
Prompt caching with cached input billing
Speculative decoding for faster token generation
Prompt caching for context reuse
Real-time prompt processing while typing
Concurrent request batching
Bit-for-bit identical output verification
Multi-GPU and multi-backend support (CUDA, Vulkan)
Vision model support with image encoding
Mixture-of-experts model support
RAD-1 automatic draft token generation
Cross-platform support for Windows and Linux
What makes it different
Prompts and model weights verified by clients with cryptographic attestation before any data is sent
Operator-blind inference: data remains unreadable to the company and hosting provider
Flexible deployment: choice of hosted cloud, rented VMs, or licensed stack on own infrastructure
Open-weight models only, no closed-source frontier models
Full cryptographic verification stack including open-source tools
Exact deterministic inference: all default optimizations preserve bit-for-bit identical output compared to plain decoding
4.5× faster performance on editing tasks on small models compared to llama.cpp
Verifiable reproducibility through receipt system that records all request parameters
Nine optimization techniques enabled by default with ability to disable each individually
No dependency on draft models or external selections for speculative decoding
Confidential AI
Kinetic