BananaMind-2-Medium-Chat-ONNX

BananaMind-2-Medium-Chat-ONNX

ONNX export of BananaMind/BananaMind-2-Medium-Chat, with KV cache, for Transformers.js and ONNX Runtime.

The weights were remapped into an equivalent Qwen3 layout (RoPE pair permutation, embedding scale folded in), so no custom code is needed. The fp32 export matches the original model's greedy output exactly.

Top-1 agreement and mean per-token KL(fp32 ‖ variant) are measured against the fp32 model over 2,000 WikiText-2 prompts (225,265 tokens, up to 128 tokens each).

fp16 is effectively lossless and q8 is close to the original. q4 is heavily degraded for a model this small and is not recommended.

import { pipeline } from "@huggingface/transformers";

const generator = await pipeline("text-generation", "BananaMind/BananaMind-2-Medium-Chat-ONNX", { dtype: "fp16" });
const out = await generator([{ role: "user", content: "What is the capital of France?" }], { max_new_tokens: 128 });
console.log(out[0].generated_text.at(-1).content);