
TensorFold
LLM Inference Engine for Metal, CUDA and Vulkan.
The Lens
By Erik Loyd, SaaS CEO and former COO/CFO of an AWS Premier Partner.
Updated Nov 2026
TensorFold runs AI language models on your own Mac or NVIDIA machine from a single Zig binary, no Python needed to serve. Pull a model, start the server, and any app that speaks the OpenAI or Anthropic API can use it. The engine, a beta Mac app and the StarLight fleet manager are all Apache-2.0 and free.
Setup is two commands through Homebrew on Apple Silicon, or a signed Linux archive on NVIDIA with driver R580 or newer. Plan around the model table: each model is qualified on specific hardware, often one M5 Mac or one GB10 (NVIDIA's DGX Spark), and the README says plainly where nothing was measured.
Verdict: impressive pace, narrow qualification. ggml-org/llama.cpp and ollama/ollama run on far more hardware, ml-explore/mlx-lm is the safer Mac pick, and vllm-project/vllm is the proven NVIDIA server. Solo: free, worth a weekend on a recent M-series Mac or DGX Spark. Small teams: free, but pin a version and test your exact model. Larger teams: StarLight adds SSO, SCIM and budgets for free; support means sponsoring the project.
The catch: the Zig engine replaced a Python one this year, one maintainer writes most of the code, and releases land every day or two.
Free vs Self-Hosted vs Paid
fully freeFree: Everything. The engine, the beta Mac desktop app and TensorFold StarLight (the multi-machine control plane with OIDC/SAML sign-in, SCIM, per-project model access and token budgets) are all Apache-2.0. The project's enterprise page says StarLight is free to use in full, including for enterprise deployments.
Self-hosted: This is the only way to run it. You supply the hardware: an Apple Silicon Mac, or a Linux (or Windows, built from source) machine with an NVIDIA GPU and driver R580 or newer. Model support is qualified per machine, so check the README's model table before buying hardware for it.
Paid: No paid tier or hosted service. Direct support, deployment help and long-term maintenance are offered only through donations, sponsorship or project backing, scoped case by case with no published prices.
Free all the way up, including the enterprise fleet manager. The real cost is your hardware and the time to track a project that ships every day or two.
What to do by team size
- Solo
- free, if you own a recent Apple Silicon Mac or NVIDIA GPU
- Small team
- free, pin a release and test your exact model first
- Medium team
- free, but llama.cpp or vLLM are safer production bets today
- Large team
- free StarLight for SSO and budgets, sponsor the project if you need support
Get tools like this every Wednesday
One featured tool, three on the radar. No fluff.
Similar Tools

Get up and running with Kimi-K2.5, GLM-5, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.

Open-source AI engine, run any model locally

High-throughput LLM inference and serving engine

SDK and proxy to call 100+ LLM APIs in OpenAI format

The Modular Platform (includes MAX & Mojo)

The fastest local AI engine for Apple Silicon. 4.2x faster than Ollama, 0.08s cached TTFT, 100% tool calling. 17 tool parsers, prompt cache, reasoning separation, cloud routing. Drop-in OpenAI replacement. Works with Claude Code, Cursor, Aider.
A low score is not a verdict on quality. Young and niche tools start low by design. How we calculate scores
License: Apache License 2.0
Apache-2.0 engine and StarLight; model weights carry their own licenses.
Commercial use: ✓ Yes
About
- Owner
- Ash (User)
- Stars
- 2,270
- Forks
- 525
Explore Further
More tools in the directory
rea
Reverse engineer anything with agents, from app behavior down to native binaries.
462.8k ★openclaw
Your own personal AI assistant. Any OS. Any Platform. The lobster way. 🦞
394.6k ★everything-claude-code
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
304.0k ★