Inference engine · Rust · NVIDIA Blackwell
Frontier open models on hardware you own.
Mouchak builds TensorHive, a memory-safe, Rust-native LLM inference engine for NVIDIA's desktop Blackwell systems: DGX Spark today, RTX PRO next. Teams run large open models privately, on their own desks, without sending data to anyone.
Talk to usTensorHive
TensorHive is our inference engine, written in Rust from the host runtime down to the GPU kernels, built on NVIDIA's cuda-oxide compiler. Every kernel is checked against a reference implementation before it ships.
Memory-safe end to end
Host, scheduler, network transport and GPU kernels in Rust. One toolchain, one binary, no Python in the serving path.
Built for GB10
Kernels tuned to DGX Spark's unified memory, and a custom RDMA transport over ConnectX-7 for multi-node serving.
Verified, not assumed
Outputs are compared step by step against a reference run. A change lands only with evidence that it is correct.
Private by design
Runs on-premises behind an OpenAI-compatible API. Your prompts and data never leave your machines.
Progress
Measured on two NVIDIA DGX Spark systems connected by ConnectX-7, serving GLM-5.3-Flash, a 321-billion-parameter mixture-of-experts model.
parameter model served across two DGX Sparks, with every reply matching the reference run over a 2,032-step test.
median latency for a 16 KiB exchange between the two Sparks on our RDMA transport, about 3× lower than the baseline engine's.
of GB10's 273 GB/s memory bandwidth reached by our Rust GPU kernels, with bit-exact results.
Numbers from our October 2026 measurements. Next: single-stream decode below 24 ms per token.
Team
Subhrajit Makur
Co-founder · Engineering
Leads the engine: GPU kernels, multi-node transport and serving.
Jayati Bairagi
Co-founder · Business & Operations
Leads partnerships, customers, finance and operations.
Contact
Running DGX Spark or RTX PRO hardware and want private inference? We'd like to hear from you.