Learn/Mindcandle

candle: Hugging Face ML Framework in the Remade Graph

Hugging Face's candle is the minimalist Rust ML framework Remade with Rust forks and pins so on-device Mind crates can share one tensor runtime.

Signed by M·
Painted small workshop lit by one calm candle flame amid soft mesh sparks
What is candle?

candle is Hugging Face's minimalist ML framework for Rust. Remade with Rust maintains a fork under the org so Mind crates can pin a shared tensor runtime: Remade did not create candle.

Hugging Face's minimalist ML framework for Rust: forked into the Remade graph, not invented here.

candle is not a Remade-with-Rust original. It is Hugging Face's candle: a small, pure-Rust ML framework for tensors and inference. Remade with Rust keeps a fork under the org so Mind crates can pin one shared runtime instead of each pulling an unconstrained upstream. This article is the builder brief: what candle is, why the fork exists, how it sits in the constellation, and how it deploys toward MATA.NETWORK.

For the portfolio thesis, read The Importance of Remade With Rust. For the map, open Vision. Filter Learn by Mind to see sibling field notes.

What candle Is For

Problem it inherits

On-device and edge products still lean on heavy Python ML runtimes for inference in places customers never see. Those stacks decide your CVE inbox, your shipping weight, and whether wasm32 or neighborhood hardware are even reachable. Remade Mind crates need a Rust-native tensor path instead.

What Remade does with it

Remade does not rewrite candle from scratch. The house:

  1. Forks and pins Hugging Face candle under Remade-With-Rust/candle.
  2. Consumes that pin from crates such as ffai, diana, argus, and mercury.
  3. Keeps attribution clear: Hugging Face owns the design and the upstream; Remade owns the fork policy and the products built on top.

That is how rebuilding the foundation of the internet stays honest: Remade rebuilds what it remakes, and names what it only carries.

Who should adopt it first

Teams already building on Remade Mind crates, or anyone who wants one Rust ML runtime shared across detection, captioning, and speech without dragging a Python serving stack onto the node.

How It Interacts With the Constellation

Inside Mind

Related Learn guides: ffai, diana, mercury, tokenizers. Those are Remade (or Remade-aligned) products that run on candle. Treat candle as shared substrate, not as a Remade SKU.

Across clusters

  • Media feeds bytes into Mind engines and Edge sensors.
  • Memory stores indexes and replicas beside the user (SpaceDB, rusty_alloc).
  • Identity gates who may call your API (mID, Sovereign ID).
  • Trust keeps the dependency graph honest (Deputy) before you place a binary on hardware you do not babysit.

That graph is the practical answer to why Digital Freedom needs Rust: ownership fails if any layer still smashes the stack or phones a landlord by default.

Business Opportunities and Implementations

Sell Remade Mind appliances and kits that happen to run on candle. Customers pay for privacy, uptime, and memory-safe product code: not for a claim that Remade invented Hugging Face's framework.

Concrete implementation patterns

  1. Library embed: depend on the house-pinned candle behind a stable facade in your Rust service; keep HTTP/JSON at the edge.
  2. Appliance SKU: ship a NUC or ESP-host kit where Remade Mind crates (detection, captioning, ASR) are the differentiator and candle is the disclosed runtime.
  3. VPC / on-prem mirror: same binary customers run in their network, with candle provenance documented for NIST SSDF-style reviews.
  4. Mesh preview: integrate now against Preview hosting; keep disco placement as a config switch on MATA.NETWORK.

What not to sell

Do not market candle as a Remade original or as a purple SaaS wrapper around Python. Credit Hugging Face for the framework; warranty the Remade products that compose on top of the fork.

Deploying Onto MATA.NETWORK

MATA.NETWORK is the distributed-cloud home for this portfolio: identity, storage, and mesh seams you can build against while hosting is still Preview. A candle-backed Mind binary should land as:

  1. A placeable artifact (native service, wasm module, or host process beside firmware).
  2. Capability-gated with mID so strangers cannot invoke it on a neighborhood node.
  3. State nearby in SpaceDB when the product produces user data.
  4. Supply-chain scanned with Deputy before disco place: including the candle pin itself.

External references: Hugging Face candle, the house fork, mata.network, and NIST SSDF. Internal next steps: Vision, Learn filter Mind, and the related articles linked above.

candle is an upstream dependency with a clear fork story: not a slide claiming Remade invented the flame.

FAQ

Quick answers for builders evaluating this technology.

Did Remade with Rust create candle?

No. candle is developed by Hugging Face. Remade with Rust keeps a fork at github.com/Remade-With-Rust/candle so portfolio crates can depend on a pinned, house-aligned copy.

What does candle replace in a Remade stack?

It is the Rust tensor / inference runtime Mind crates build on, so product code can avoid hauling a heavy Python ML stack onto edge and mesh nodes.

Which Learn cluster does candle belong to?

Mind. Use the Learn filters to browse sibling crates that run on top of candle (ffai, diana, argus, mercury, tokenizers).

How does candle help a product team commercially?

It is an upstream dependency you can warranty by pin and fork policy: one Rust ML runtime across captioning, detection, and ASR crates, with a clear provenance story for procurement.

Where is the source for candle?

Upstream: https://github.com/huggingface/candle. House fork used by Remade with Rust: https://github.com/Remade-With-Rust/candle.

← All articles