celsowm/lfm2.5-350m-wasm-q4_0

LFM 2.5-350M — 4-bit (Q4_0) for the pure-WASM engine

A 4-bit quantized build of LiquidAI's LFM 2.5-350M text model, packaged for the lfm-wasm inference engine — a pure WebAssembly runtime with no native code, no server, and no network at inference time. Everything runs in the browser.

⚠️ Custom format. These files are not GGUF, safetensors, or PyTorch weights. They are consumed directly by the companion lfm-wasm engine in this repo. Do not load them with transformers / llama.cpp.

🌐 Live demo: https://celsowm.github.io/lfm-wasm/ — the model runs entirely in your browser (fetched from this repo at runtime).

Files

File	Size	Description
model_q4_0.bin	~272 MB	Quantized weights + model header, Q4_0 (4-bit) packed.
vocab.json	~0.7 MB	Byte-level BPE vocabulary: a JSON array of token strings.
merges.json	~1.1 MB	BPE merge rules: a JSON array of [left, right] pairs.
README.md	—	This card.

Using it

The files are fetched by the web playground at runtime:

const [model, vocab, merges] = await Promise.all([
  fetch("/model_q4_0.bin"),
  fetch("/vocab.json"),
  fetch("/merges.json"),
]);
const engine = new LFM2Engine(
  new Uint8Array(await model.arrayBuffer()),
  await vocab.text(),
  await merges.text()
);

Weights are Q4_0 (4-bit block quantization, 32-fp16-scale per 32 weights) with Q8_0 attention projections; inference uses WASM SIMD gemm/gemv kernels. See the source repo for the full engine, tokenizer, and ChatML prompt template.

Provenance

• Base model: LFM 2.5-350M by LiquidAI.
• Quantized locally with the repo's quantize tool (cargo run --bin quantize) from the original bf16 checkpoint.
• No fine-tuning or weight modification beyond quantization.

License

Inherited from the base LFM 2.5 release. See LiquidAI's license terms for LFM 2.5 before use in production.