TensorFold Labs Ltd
TensorFold builds open-source local-inference software that runs language models quickly on hardware customers own, with native Metal and CUDA engines for Apple Silicon and NVIDIA GPUs. Its enterprise StarLight product adds identity, deployment, policy, routing, budget and usage controls around private AI fleets.
At a Glance
- Individual developers and local-AI enthusiasts
- Apple Silicon Mac owners
- NVIDIA GPU, DGX Spark and GB10 operators
- Coding-agent and developer-tool users
- +2 more
AI Tools by TensorFold Labs Ltd
(1)
TensorFold
Local LLM Inference Engine
Discussions
No discussions yet
Be the first to start a discussion about TensorFold Labs Ltd
Latest News
TensorFold Labs Ltd incorporated in England and Wales as a private limited software-development company
TensorFold 1.2.1 added CUDA models/backends, 1M-token Qwen context, Gemma 4 drafting, MOSS-Image support and further GB10/Arc performance improvements
TensorFold 1.2.0 added MiniMax-M3.1-Flash on Metal, new CUDA model/checkpoint support and Intel Arc Pro B70 support
TensorFold 1.1.9 improved DGX Spark prompt performance and Flash Next on Macs while retaining output-token equivalence
Products & Services
Open-source Zig-based runtime and HTTP server for serving local language models on Apple Silicon/Metal and NVIDIA/CUDA (with additional platform work). It reads checkpoints directly, provides OpenAI-compatible chat/completions and Responses routes plus Anthropic Messages, streaming, tool calls, tokenization, health and metrics endpoints.
Apple-Silicon macOS desktop workspace with the TensorFold engine built in, model downloads, local API, coding/document workers, tool-use loops, streaming responses, live diffs, approval gates and local-first storage.
Enterprise AI control plane for private TensorFold fleets. It provides one governed OpenAI- and Anthropic-compatible endpoint, SSO, SCIM, project permissions, model deployment and placement, fleet/node operations, budgets, routing, metering, audit and optional approved cloud-provider routes.
Supporting enterprise components: Connect provides sign-in and a local endpoint for desktop tools, while Starlight Node runs and supervises TensorFold engines on macOS, Linux and Windows machines.
Market Position
TensorFold positions itself as a high-performance, local-first inference engine optimized for low latency and exact output preservation, rather than a model provider. Public comparisons place it against vLLM, SGLang, Ollama and llama.cpp; its claimed advantage is hardware-specific kernels and exact drafted decoding for single-user and small-concurrency workloads, while vLLM is described as stronger for very high-concurrency continuous batching. StarLight extends the position from a runtime into a governed enterprise control plane.
Leadership
Founders
Ashley (Ash) Steven Hart
British AI systems architect and software developer based in England. His public profile says he builds and tests local AI systems on Apple Silicon and NVIDIA hardware, covering MLX, CUDA, mixed-hardware inference, local agents and runtime security; he has also built MCDMA (Metal-CUDA Direct Memory Access) and Imprint. Companies House lists him as the company's sole active director, appointed 2 November 2026.
Executive Team
Ashley (Ash) Steven Hart
Founder and Director
Independent AI systems architect and developer focused on MLX, CUDA, local inference, mixed-hardware systems, local agents and runtime security; creator of TensorFold and related projects including MCDMA and Imprint.
Founding Story
TensorFold began as Ash Hart's weekend experiment to make Qwen 27B faster. The initial vision was to make local inference substantially faster without changing model outputs, by writing hardware-specific kernels and using exact/speculative drafted decoding; it subsequently grew into an open-source engine and an enterprise control-plane effort.
Business Model
Revenue Model
The core TensorFold engine and Mac app are distributed as open-source/open-compatible software (Apache-2.0 is stated for current releases). The commercial path is enterprise deployment, support and StarLight's organization control-plane capabilities; the site invites enterprise deployment/support and investment/partnership enquiries.
Target Markets
- Individual developers and local-AI enthusiasts
- Apple Silicon Mac owners
- NVIDIA GPU, DGX Spark and GB10 operators
- Coding-agent and developer-tool users
- Organizations operating private AI fleets
- Teams needing governed local and approved cloud model access
- Private local chat and LLM serving on Apple Silicon Macs
- Low-latency inference on NVIDIA GPUs and DGX Spark/GB10 systems
- Coding agents, SDKs and applications using OpenAI- or Anthropic-compatible APIs
- Multimodal and long-context model serving where supported by the model/backend
- Multi-machine/tensor-parallel deployments across connected Macs or GPUs
- Enterprise AI fleets requiring private deployment, governance, identity, budgets and usage accounting