Lanna Script and NLP/LLM

Introduction to Kam Mueang (Northern Thai Language)

Northern Thai, also known as Kam Mueang, is a Southwestern Tai language spoken by approximately six million people, primarily in Northern Thailand. Traditionally written in the Tai Tham script, it is nowadays most commonly transcribed using the standard Thai script. However, this modern transcription often fails to accurately capture the language’s unique six-tone system. While closely related to Standard Thai, Kam Mueang retains a distinct vocabulary and phonology deeply rooted in the history of the Lan Na Kingdom. Today, the language faces increasing pressure from “Thaification,” leading many native speakers to code-switch between Kam Mueang and Standard Thai, especially in digital and urban spaces. ¹

Introduction to Lanna Script

The Lanna script, also known as the Tai Tham script, is an abugida used to write several Tai languages, including Kam Mueang, Tai Lue, and Khuen. It is believed to have originated from the Old Mon script and was introduced to the Lan Na Kingdom in the 13th century. The script consists of 44 consonants, 15 vowels, and several tone markers. ²

Why Lanna Script and NLP/LLM?

As a native speaker of Kam Mueang, I have always been fascinated by my language and its rich history. Recently, I have become particularly interested in the potential of Natural Language Processing (NLP) and Large Language Models (LLMs) to help preserve and promote this low-resource language. This blog serves as a space where I can share my ongoing research and thoughts on these topics.

Being able to type the Lanna script natively across digital devices is a monumental step forward for our culture. I want to express my deep gratitude to the Unicode Consortium for making this possible, and for their years of hard work dating back to the early 2000s.

LLM Capabilities on Lanna Script

I recently read in the Tai Tham Unicode Facebook group that Gemini 3 can now generate and understand the Lanna script. Intrigued by this capability, I decided to investigate it further. I began testing multiple LLMs currently available online, spanning both open-source and closed-source models. Here is the complete list of LLMs I evaluated:

Open Source LLMs (Available on HuggingFace)

• Qwen3.5 397B-A17B
• DeepSeek-V3.2-Speciale
• Mistral Large 3-2512
• Llama 4 Maverick-17B-128E-Instruct
• Llama 4 Scout-17B-16E-Instruct
• Qwen3 235B-A22B-Thinking-2507
• DeepSeek R1
• Kimi-K2-Thinking
• Kimi-K2.5
• MiniMax M2.1
• GLM-4.7
• Deepseek V3.2

Closed Source LLMs (Openrouter API)

• Qwen3.5 Plus 2026-02-15
• GPT-5.2
• Mistral Medium 3.1
• Claude Sonnet 4.5
• Claude Opus 4.5
• Gemini 3 Flash
• Gemini 3 Pro

Results

In this section, I will share the results of evaluating these LLMs on Lanna translation tasks. Please note that not all LLMs were evaluated using the exact same prompt and input text; however, the results still provide a robust general idea of each model’s performance. I have divided the evaluation into two distinct parts: Natural Language Understanding and Natural Language Generation.

1. Natural Language Understanding

Here is the prompt I used for the test:

Prompt: Can you translate this passage to english: [LANNA_TEXT]

The source material was drawn from the Lao Geography book (1896) ³ (Note: “Lao” was historically used by Siam to refer to the Lan Na people during the 19th century when Lan Na was a vassal state).

Here are my input texts:

Here is another example:

Also, there is another input text. I manually translated the question from the Global MMLU Lite Dataset ⁴.

LLM Performance Evaluation: Tai Tham (Lanna) Translation

Lanna LLM Scoring Framework

1. Translation Quality Score

2. Generation Time Score

2. Natural Language Generation

For the Natural Language Generation test, I primarily worked with Gemini 3 Flash (currently, I see no distinguishable difference in writing quality between Gemini 3 Flash and Gemini 3 Pro). The Gemini 3 Series models are undeniably fast, cheap, and highly proficient at Lanna string manipulation. While Claude Opus 4.5/4.6 are also remarkably good at Lanna, they remain prohibitively expensive. To put it in perspective, Gemini 3 Flash is the most economical choice at around $0.01 per question, while Gemini 3 Pro costs roughly $0.15 per question. In stark contrast, Claude Opus 4.5/4.6 costs over $1.00 per question. I gave the models basic instructions to translate an English phrase into Lanna.

I chose the famous quote from John F. Kennedy:

We choose to go to the Moon not because it is easy, but because it is hard. – John F. Kennedy

As seen above, even with relatively simple vocabulary, Gemini still occasionally introduces typos and spelling inconsistencies. How should we tackle this limitation moving forward? That remains an open question.

Why are paid models better at Lanna?

In this section, I want to discuss a few interesting observations from our evaluation:

1. It is abundantly clear that open-source models currently lag far behind proprietary, paid models when it comes to the Lanna script.
2. Gemini stands out as both economically viable and lightning-fast, demonstrating a remarkably high understanding of the script.
3. Claude also possesses a strong grasp of Lanna, but its high inference cost makes it less accessible.

The big question is: WHY are paid models so much better at Lanna? I will focus my theory on Gemini 3.

The Lanna script features a deeply complex orthographic structure, completely unlike English and significantly more intricate than modern Thai. Furthermore, it is a script used almost exclusively by a minority group in Thailand and lacks official government backing or standardized educational support. Consequently, there is an extreme scarcity of digitized Lanna data available online. I estimate that the total volume of “ready-to-use” digital Lanna text might only range between 20,000 to 200,000 words. This data drought is exacerbated by the fact that most people still write Lanna strictly on physical paper or rely heavily on Thai-based fonts, simply because standard Tai Tham Unicode is not yet fully supported by ubiquitous software like Microsoft Office.

An interesting recent research paper might help explain Gemini’s dominance in this area. Titled “Assessing Thai Dialect Performance in LLMs with Automatic Benchmarks and Human Evaluation” ⁵, the study explores how effectively modern LLMs process localized Thai dialects.

Here are two key findings from the paper:

1. Gemini performs best across all local Thai dialects when compared to its peers. Human evaluators consistently preferred Gemini’s outputs across the board.
2. No open-source models (including Llama and Typhoon—a Thai-centric model developed by SCB 10X) performed adequately in these evaluations.

These findings perfectly align with my own experiments. The data overwhelmingly suggests that Gemini currently reigns supreme in Lanna language tasks.

Final Thoughts

1. Current open-source models still lack the fundamental capability to process and understand the Lanna script.
2. We urgently need more comprehensive, standardized benchmarks to accurately assess real-world model capabilities in this language.
3. It may already be viable to build specialized applications utilizing Lanna text on top of the Gemini API.
4. Translating from Lanna to higher-resource languages (like English or Standard Thai) is largely possible today.
5. Conversely, translating from higher-resource languages into Lanna still requires heavy human intervention, although semi-automated, “human-in-the-loop” workflows are becoming feasible.
6. LLMs suffer from a severe lack of Northern Thai vocabulary knowledge. Models frequently hallucinate by substituting Central Thai words written in Lanna script, rather than using authentic Northern Thai vocabulary (a phenomenon colloquially known as ไทยปะแล๊ดเมือง).
7. Systematically digitizing the Lanna–Thai dictionary and pursuing targeted fine-tuning could drastically alleviate this issue. For instance, the renowned “พจนานุกรมล้านนา-ไทย ฉบับ แม่ฟ้าหลวง” (Mae Fah Luang Lanna-Thai Dictionary) contains roughly 22,000 entries. In contrast, the current Lanna database on Wiktionary holds only about 1,500 words. Injecting even a few thousand well-curated entries into training pipelines could teach LLMs proper orthography—especially for complex loanwords from Pali, Sanskrit, and English, whose Lanna spellings deviate wildly from native Northern Thai conventions.

Future Work

Moving forward, I plan to focus my efforts on two primary initiatives:

1. Developing formal, comprehensive evaluations and benchmarks for LLMs operating on the Lanna script.
2. Contributing to the active digitization of the Lanna–Thai dictionary to expand the open-source corpus.

While the gap between massive, proprietary AI models and open-source alternatives remains vast for low-resource languages like Lanna, it is genuinely exciting to see tools like Gemini finally making our script digitally functional.

Thank you so much for reading! I hope you found this deep dive into the intersection of Kam Mueang heritage and modern AI technology as fascinating as I did.