THaLLE
Kasikornbank tech arm's open-weight finance LLMs, now in Thai as well as English
Overview
THaLLE (Text Hyperlocally Augmented Large Language Extension) is a family of finance-specialised small language models from KBTG Labs, part of Kasikornbank's technology group. Version 0.1 (2024) is a Qwen2-7B-Instruct fine-tune on CFA mock exams; version 0.2 (December 2025) merges Qwen3-8B, ThaiLLM-8B and a finance fine-tune for Thai and English. Weights are free under Apache 2.0 on Hugging Face.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 19 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
What you can produce with THaLLE
- Open Apache 2.0 finance-tuned model weights to run on your own hardware
- Thai and English finance answers from the 0.2 model, with a switchable reasoning mode
- An official GGUF build of the 0.1 model for lightweight local inference
- Three arXiv technical reports documenting training, merging and evaluation
- Published benchmark results on Flare CFA, Thai IC and O-NET exams to reproduce
- A base for on-premise Thai financial chatbots or Q&A tools, with your own safety layer
ASEAN Perspective
THaLLE in Southeast Asia
THaLLE is narrower than the other Thai models here and more interesting for it. KBTG, the technology arm of Kasikornbank, built it for finance and evaluated it against the Stock Exchange of Thailand's Investment Consultant examination, scoring 72%, 72% and 84% across the three levels. The benchmark is regionally significant because it tests Thai regulatory and product knowledge that general models lack. It should be treated as a research artifact, not a product. The weights are available for research, but there is no commercial support. Its existence, however, proves that domain-specific Thai models can be built at a size a bank can run internally.
THaLLE is a well-documented research project rather than a product: open Apache 2.0 weights, three arXiv reports, and a 0.2 model that adds Thai and a reasoning mode to finance knowledge tested on CFA and Thai Investment Consultant exams. The best-known 72/72/84 IC result belongs to a variant KBTG never released, the exam sets are small, and every score is the lab's own.
KBTG's own tests show weak refusal of harmful prompts unless a safety instruction is added. Useful for Thai fintech and banking developers who want an on-premise starting point; not something to put in front of customers without your own evaluation and guardrails.
What we found
THaLLE (Text Hyperlocally Augmented Large Language Extension) is a family of small finance-focused language models from KBTG Labs, the research unit of Kasikorn Business-Technology Group, the technology company of Thailand's Kasikornbank. KBTG describes it as built for the bank's own needs, but it also shares the work openly: the released weights sit on Hugging Face under Apache 2.0, which permits commercial use, and each version has an arXiv technical report. There is nothing to buy, no hosted API and no support.
Three reports trace its development. The June 2024 report introduced THaLLE-0.1-7B-fa, a Qwen2-7B-Instruct model fine-tuned with LoRA on 9,426 internal CFA mock-exam questions; it scored 72% on the public Flare CFA benchmark, against 88% for GPT-4o. The November 2024 report adapted it to Thai finance using the Stock Exchange of Thailand's Investment Consultant exam. Its THaLLE-0.1-7B-ic variant scored 72%, 72% and 84% on the P1, P2 and P3 practice exams, matching GPT-4 models on P3, but that variant was never released and P3 has only 25 questions. The downloadable 0.1 checkpoint scored 64%, 52% and 60%.
The current release, THaLLE-0.2-ThaiLLM-8B-fa, appeared in December 2025: an equal-weight merge of Qwen3-8B, the Thai-pretrained ThaiLLM-8B and a finance fine-tune, with thinking and non-thinking modes and Thai as well as English. KBTG reports 84% on a 25-question Thai IC set in thinking mode, against 60% for Qwen3-8B, and 85% on Flare CFA. Its own safety table is the warning: without an explicit safety instruction the model refused only about 30% of harmful Thai prompts, and the authors recommend adding one for consumer-facing use.
Every score here is the lab's own, with no outside reproduction. For Thai fintech teams wanting an on-premise starting point, it is a credible research base, not a finished product and not a source of financial advice.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including THaLLE's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with THaLLE unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to THaLLE directly →
Spotted something out of date? Suggest an update →
Alternatives to THaLLE
More in LLMs & Chat
