chandra-twin-en-onnx

The on-device "AI twin" that answers questions on tungathurthi.com: a LoRA fine-tune of LFM2.5-230M that speaks in first person as Chandra Tungathurthi (AI consultant, Essen, Germany) and answers questions about his work, products and sovereign-AI consulting. English only.

The model is trained to answer from retrieved context injected into the system turn (### External Info: blocks) and emits plain prose only — links and actions are resolved by the consuming runtime (twin-response/v1 contract). It is served fully client-side via Transformers.js + WebGPU; conversations never leave the visitor's device.

Variants

File	Precision	Size	Notes
onnx/model_fp16.onnx	fp16	~454 MB	Recommended — matches the PyTorch model on the behavioral scorecard
onnx/model_q4.onnx	int4 (block 32)	~211 MB	Smaller and faster; measurably lossier at this model size

Exported with LiquidONNX.

Usage (Transformers.js)

import { AutoModelForCausalLM, AutoTokenizer } from "@huggingface/transformers";

const modelId = "tckb/chandra-twin-en-onnx";
const tokenizer = await AutoTokenizer.from_pretrained(modelId);
const model = await AutoModelForCausalLM.from_pretrained(modelId, {
  device: "webgpu",
  dtype: "fp16",
});

Serving parameters the model was validated with: greedy decoding, repetition_penalty: 1.05, ChatML template with the system prompt and retrieval context of the website's twin runtime.

Intended use

Built for one purpose: powering the chat on Chandra's own website. Answers may be imperfect — the authoritative source is the person: letsconnect@tungathurthi.com.