amityco/lbm-v3-2

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 28, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

amityco/lbm-v3-2 is a 27 billion parameter language model developed by amityco, fine-tuned from Qwen3.5-27B for consumer digital-twin applications in Southeast Asia. It specializes in client-instrument tasks, excelling in client pricing surveys, survey-panel replication, and churn prediction. The model demonstrates strong performance in Brand's WTP structure, HBA panel TVD, and specific use-case AUC metrics for churn and purchase.

Loading preview...

Model Overview

amityco/lbm-v3-2 is a 27 billion parameter language model, a flagship "client-instrument specialist" developed by amityco. It is fine-tuned from Qwen3.5-27B, with further stages of Supervised Fine-Tuning (SFT) on proprietary datasets (amityco/lbm-sft-v32) focusing on SEA consumer voice, public-source personas, and reviews. This model is specifically designed for consumer digital-twin applications within the Southeast Asian market.

Key Capabilities & Performance

The model demonstrates strong performance across several specialized benchmarks:

  • Brand's WTP: Achieves a structure 'r' of 0.990, a MAPE of 19.2%, and a T2B of 0.547.
  • HBA panel TVD: Scores 36.1 with a 63% top-match rate.
  • Coupon Prediction (proprietary TH retailer): Shows a Brier score of 0.469 and a meanP of 0.759.
  • Use-case Specific AUC: Excels in churn prediction with an AUC of 0.780 and purchase prediction with 0.563.
  • Bangkok Likert: Achieves an 'r' of 0.393 and MAE of 0.420.

Recommended Use Cases

Based on its specialized fine-tuning and evaluation results, amityco/lbm-v3-2 is particularly well-suited for:

  • Client pricing surveys (Brand's-style)
  • Survey-panel replication
  • Churn prediction

Limitations

Users should avoid this model for:

  • Persona-conditioned interest prediction (ov@5 score of 0.084)
  • Open-ended VW price ladders (due to a 1% monotone issue and ladder-collapse in v3.2)