Zero dependency
No Bootstrap, jQuery, React, Vue, Tailwind, tokenizer crate, ML framework, wasm-bindgen requirement, remote inference API, CDN asset, or npm runtime.
Project Website
TinyRustLM: Browser-Local Tiny Language Models in Rust and WASM
TinyRustLM is a project site for a custom Rust/WASM SLM runtime that keeps model execution local, small, auditable, and dependency-free.
Custom .slm payloads
A compact binary format with strict header, tensor directory, tokenizer metadata, quantization fields, and checksum validation.
Browser-first verification
Loopback server, static app contract harness, WASM ABI smoke tests, browser smoke modes, and deterministic output checks.
Adapters and assembly
ADP1 raw deltas, ASP1 sparse deltas, ALR1 low-rank deltas, selector registries, module-plan receipts, and apply-before-generate gates.
01
What TinyRustLM is
TinyRustLM is a browser-local small language model runtime written for Rust and WebAssembly. The project is not a generic AI website and not a hosted chatbot. It is an implementation-centered site for explaining how a custom runtime loads compact .slm model artifacts, validates their shape and checksums, executes generation inside the browser sandbox, and exposes diagnostics without remote inference.
- Custom Rust/WASM runtime
- No remote inference API
- No JavaScript framework
- No ML framework runtime
- Local .slm artifacts and manifests
- Deterministic smoke-test posture
02
The core constraint
The design centers on a strict browser selector model byte budget of 33,554,432 bytes. This pressure forces the site and the runtime to explain why model shape, vocabulary size, quantization, scratch buffers, and adapter payloads are first-class product concerns rather than hidden implementation details.
- 33.5 MB selector model envelope
- q8_0 and q4_0 data paths
- Bounded prompt and output buffers
- Reusable runtime scratch
- Explicit failure states
03
Why it matters
TinyRustLM treats tiny models as local compute units: small enough to fetch, parse, validate, and run in constrained browser memory, but structured enough to test tokenization, quantization, adapter application, provenance, and deterministic generation. The project is most useful as a transparent engineering baseline for local AI experiments, not as a claim that microscopic models equal cloud-scale assistants.
- Local privacy by default
- Supply-chain surface reduction
- Inspectable artifacts
- Fast failure and recovery
- Model provenance before model claims
Plain PHP deployment notes
No WordPress bootstrap, theme system, database requirement, composer package, npm build, CDN script, Bootstrap class dependency, or jQuery dependency.
Routes are handled by index.php and server rewrites. Core content lives in data/pages.php. HTML and Markdown share the same page data.
Upload the package to the subdomain root, point Apache/Nginx to the folder, and keep .htaccess enabled for clean routes on Apache.