FinnAI SLM v4

On-device language model for Indian bank SMS → JSON and personal finance coaching.

Use this model when you need private, phone-friendly understanding of Indian banking SMS and short finance answers — without sending user messages to the cloud.


Model	`finndot/finnai-slm-v4`
On-device INT4	`finndot/finnai-slm-v4-litertlm`
Dataset	`finndot/finnai-slm-data-v4`
Base	`Qwen/Qwen3-1.7B`
License	Apache 2.0
App	FinnDot

FinnAI model family

Version	Weights	On-device LiteRT	Data
v4 (latest)	`finnai-slm-v4`	`finnai-slm-v4-litertlm`	`finnai-slm-data-v4`
v3	`finnai-slm-v3`	`finnai-slm-v3-litertlm`	`finnai-slm-data-v3`
v2	`finnai-slm-v2`	—	`finnai-slm-data`

Prefer v4 for new apps and fine-tunes: broader SMS coverage, Ask / Learn / Analytics chat aligned with FinnDot.

What it helps with

Capability	How it helps
SMS → JSON	Turns UPI / NEFT / IMPS / card / salary / ATM SMS into `{amount, merchant, type, account, balance, category}`
Noise refusal	Returns `{}` for OTP, promo, KYC — avoids fake transactions
Ask	Short answers grounded in a spending ledger
Learn	Tutor-style explanations (SIP, EMI, tax, budgeting)
Analytics	Spending tips without inventing numbers
On-device	INT4 LiteRT build runs on Android; user SMS stays on the phone

Languages: English, Hindi, Hinglish, and other Indic bank SMS styles in the training mix.

Quick start

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "finndot/finnai-slm-v4"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype="bfloat16", device_map="auto"
)

messages = [
    {
        "role": "system",
        "content": (
            "Extract transaction details from this SMS as JSON: "
            "{amount, merchant, type, account, balance, category}. "
            "type is one of: INCOME, EXPENSE, CREDIT, TRANSFER, INVESTMENT. "
            "If this is not a transaction (OTP, promo, KYC), return {}."
        ),
    },
    {
        "role": "user",
        "content": "HDFC Bank: Rs.499.00 debited from A/c XX1234 to SWIGGY via UPI. Avl Bal Rs.15000.50",
    },
]
text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

LoRA adapter: adapter/ in this repo (base: finndot/finnai-slm-v3).

Quality snapshot (held-out test)

Metric	FinnAI v4	Untuned Qwen3-1.7B
SMS R-EM %	94.87	29.39
Amount EM %	99.7	78.75
JSON valid %	100.0	100.0
Chat groundedness %	80.77	47.44

Full report: eval/v20261005-finnai-slm-v4.md.

On-device file

For Android / LiteRT-LM, use finndot/finnai-slm-v4-litertlm (~929 MB .litertlm).

Limitations

Best on Indian banking SMS; weaker on unfamiliar foreign formats.
Merchant names can still be hard; bank-specific regex parsers may win when available.
1.7B — specialized for SMS + finance chat, not a general assistant.
Thinking / long chain-of-thought is disabled for on-device use.

Citation

@misc{finnai-slm-v4,
  title={FinnAI SLM v4: On-device Indian bank SMS parsing and finance coaching},
  author={FinnDot Team},
  year={2026},
  url={https://huggingface.co/finndot/finnai-slm-v4}
}

License

Apache 2.0 — same as Qwen3-1.7B.

finndot/finnai-slm-v4