finndot/finnai-slm-v4

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 6, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The finndot/finnai-slm-v4 is a 1.7 billion parameter language model developed by FinnDot, fine-tuned from Qwen/Qwen3-1.7B with a 32768 token context length. It is specifically optimized for on-device processing of Indian banking SMS messages, extracting structured JSON data, and providing ledger-grounded finance coaching. This model excels at tasks like transaction detail extraction from SMS and offers finance-related chat capabilities in English, Hindi, and other Indic languages.

Loading preview...

FinnAI SLM v4: Specialized On-Device Finance AI

FinnAI SLM v4 is a 1.7 billion parameter language model, continually fine-tuned by FinnDot from Qwen3-1.7B, designed for on-device financial applications. Its primary function is to process Indian banking SMS messages, extracting structured JSON data for transaction details (amount, merchant, type, account, balance, category). The model also provides ledger-grounded finance coaching, offering short answers, tutor-style explanations (e.g., SIP, EMI, tax), and spending pattern analytics tips.

Key Capabilities

  • SMS to JSON Extraction: Accurately converts Indian bank SMS (UPI, NEFT, IMPS, CC, salary, ATM) into structured JSON, returning an empty object for non-transactional messages like OTPs or promotions.
  • Multilingual Support: Handles English, Hindi, Hinglish, Tamil, Telugu, Marathi, and Bengali, reflecting the diverse linguistic landscape of Indian banking.
  • On-Device Optimization: Built for mobile environments, with a LiteRT-LM INT4 version available for efficient, private processing directly on user devices.
  • Finance Coaching: Engages in ledger-grounded chat for financial queries, learning, and analytics, without inventing numbers.

Performance Highlights

FinnAI SLM v4 demonstrates significant improvements over its base model and other alternatives:

  • SMS Record Exact-Match: Achieves 94.87% on a synthetic test set (n=2028), substantially outperforming Qwen3-1.7B (29.39%) and Qwen2.5-1.5B (49.9%).
  • JSON Validity: Maintains 100% JSON validity for extracted data.
  • Chat Groundedness: Scores 80.77% on ledger-grounded finance chat evaluations (n=78), compared to 47.44% for Qwen3-1.7B.

Good For

  • Developers building on-device/mobile finance assistants requiring robust Indian bank SMS parsing.
  • Applications needing structured data extraction from diverse Indian banking SMS formats.
  • Creating private, ledger-grounded finance coaching or analytics tools for the Indian market.
  • Continual fine-tuning from a strong, specialized SMS+finance base.

Finetunes
2 models