Elffuss-LM-1.2B-ONNX

Elffuss LM — LFM2.5-1.2B healed for in-browser agentic work (ONNX q4)

The browser brain of the Elffuss family (Elffuss Claw · Elffuss Code · Elffuss) — LFM2.5-1.2B-Instruct (© Liquid AI, LFM Open License) with a small agentic alignment retrain, packaged as ONNX q4 (~850 MB) for transformers.js / WebGPU. Runs fully client-side.

What the retrain does

Aligned to Elffuss's three-mode contract — held-out validated (prompts not in training):

1. App requests → complete self-contained HTML documents (inline CSS/JS, dark, <canvas> for visual apps).
2. System actions → clean ```tool JSON: asked "recuérdame en 30 minutos llamar a mamá" → {"tool":"tasks.add","args":{"inMinutes":30,"prompt":"llamar a mamá"}}. The stock model tended to ramble instead of acting; the retrain fixes exactly this.
3. Otherwise → short warm replies.

Engineering notes (honest)

• Training: LoRA on the language model, merged into the weights. Fidelity verified vs the fp32 torch model: argmax match, cosine 1.0 on held-out prompts.
• Packaging: optimum has no lfm2 ONNX export yet, so this repo uses the official ONNX graph (byte-identical ops/encodings — including the quantized embedding, proven in-browser) with only the retrained linear weights' MatMulNBits payloads (92 layers × quant/scales/zp, block 32, asymmetric) replaced, then verified numerically.
• Use exactly like the official ONNX repo: pipeline('text-generation', 'KikoCis/Elffuss-LM-1.2B-ONNX', { device:'webgpu', dtype:'q4' }).

Base model and license are Liquid AI's. This is a derivative under the LFM Open License 1.0.