On September 18, we hosted the very first Typhoon Community Meetup! 🎉 It was a great opportunity for real Typhoon users to meet, connect, and share their experiences in person.
If you couldn’t make it, don’t worry—this blog brings you a quick recap of the key highlights, covering three main parts:
-
Updates on Typhoon’s progress and roadmap, directly from the team
-
Real-world use cases shared by the community on how Typhoon is applied in businesses
-
Relaxed networking for onsite attendees, sparking new connections and ideas
What’s New with Typhoon Models
Model Roadmap
In 2025, the Typhoon team is ramping up development and releasing more models than before, with several clear shifts from 2024:
Focus on Real-World Applications
This year’s models are designed with sharper use cases in mind, and are production-ready—easy to deploy and available via API.
Multimodality
We first released models supporting images and audio in late 2024, but those were primarily Research Previews. In 2025, we’re extending this toward practical use cases—models that are accurate, fast, and cost-efficient.
Purpose-Built for Real Use Cases; Small, Accurate, Efficient, Fast
Typhoon’s strength lies in Thai language accuracy, on par with or exceeding leading models in some metrics. But what really stands out is its compact size, enabling fast inference and significantly lower costs.
“Fast” doesn’t just mean runtime—it also reflects our small, agile team that works closely with the community, iterating quickly based on feedback. A clear example: Typhoon OCR v2 will be released soon, building directly on feedback from users of version 1.
Open Source, Open License
We’re reinforcing our commitment to open source and permissible licenses, enabling businesses to freely deploy Typhoon in production—whether self-hosted, via API, or through GPU infrastructure providers. At the meetup, we saw multiple community examples demonstrating all these approaches.
Review of 2025 Achievements
The major model releases this year clearly reflect our roadmap priorities: stronger use cases, multimodality, compact design, and cost efficiency.
Community adoption and engagement have grown significantly compared to last year:
- Nearly 20,000 unique API users
- Over 35 million API calls
- 6.5 million downloads of Typhoon models on Hugging Face
On the research side, Typhoon continues to contribute with papers published at leading conferences including ACL, Interspeech, NeurIPS, ICLR, and EMNLP.
Upcoming Models to Watch
Building on these directions, we’re actively developing several new models:
-
Typhoon OCR v2
-
Typhoon 2.5 – a text model building on 2.1 Gemma, with enhanced Agentic capabilities
-
Typhoon Isan ASR & TTS – speech recognition and synthesis for the Isan dialect
-
Typhoon Embedding – an embedding model optimized for RAG applications
Community Lightning Talks
Another highlight of the meetup was the Lightning Talks, where real users showcased how they’ve applied Typhoon in production. We had five case studies spanning enterprises, startups, government, and research.
1. Document Classification for Enterprise with Typhoon OCR
Toyota Leasing (Thailand) – Seehapong Seeharod, Pantheb Tachajarrupan, and Atiwat Tanapongworapa
The auto leasing business handles massive volumes of documents—from contracts and agreements to branch-level forms. Previously, these documents were aggregated without categorization, forcing the HQ team to manually classify them.
Toyota Leasing adopted Typhoon OCR as the core of their Document Processing Pipeline, with three critical steps:
-
Pre-Processing: Fixing issues like skewed or low-quality scans before OCR (a key factor for accuracy).
-
OCR with Typhoon: Extracting all text without filtering.
-
Post-Processing: Selecting relevant text and correcting errors using LLMs or scripts.
This pipeline now processes tens of thousands of pages monthly, significantly reducing manual workload. Pantheb emphasized that pre- and post-processing are essential for OCR to work effectively at enterprise scale.
To ensure document privacy, Toyota Leasing also used a private GPU service provided by Float16, suitable for their large document volumes. The team also shared plans to expand Typhoon’s use into other areas of the business.
2. Automatic Social Comment Responder
Pawarit Piamprecha, Tobtan.ai
Born out of the Typhoon Hackathon 2024, Tobtan.ai is a startup helping creators manage their social communities—starting with automatic comment replies.
They chose Typhoon LLM for its deep understanding of Thai and the ability to fine-tune tone and personality for each creator (e.g. playful, polite, serious). This makes replies feel natural, not templated or robotic.
Example of style-tuned comment replies
The result: creators can scale their engagement while staying authentic. Tobtan.ai is currently piloting with early adopters and open to more interested creators.
3. Typhoon LLM: Modernizing Government Workflows for Enhanced Efficiency
Pattrawut Kunwipusit and team, VAP Solutions
VAP Solutions shared their government project using Typhoon OCR and Typhoon 2.1 Gemma.
Their previous OCR (e.g. Tesseract.js) achieved only 30–60% accuracy, and did not support Thai numerals (which they had to fine-tune manually).
Switching to Typhoon OCR nearly doubled accuracy, with built-in Thai numeral support.
For text processing, they used Typhoon Text Model to summarize and categorize content, solving consistency issues caused by human reviewers. They fine-tuned with annotated data and validated with human eval, achieving reliable quality.
Crucially, the project required on-premise deployment—an area where Typhoon’s flexibility was a strong fit.
4. Typhoon OCR: Extracting & Tagging for Smarter HRM & KMS Chatbot
Natdhanai Praneenatthavee, Inteltion
Inteltion prototyped an HRM Chatbot that answers employee queries based on their internal Knowledge Base, powered by Typhoon OCR. A standout finding was Typhoon OCR’s ability to handle complex documents like charts and graphs.
AWS Architecture with Typhoon OCR Service (KMS & Chatbot Solution)
Inteltion shared their production-grade architecture on AWS:
-
AWS EKS (Elastic Kubernetes Service): Runs containerized services (Chatbot UI, KMS website, Typhoon OCR), with rolling updates via CI/CD—no need to manage the control-plane.
-
AWS ECR (Elastic Container Registry): Stores container images for Chatbot UI, KMS site, and OCR workers, integrated with CI/CD for seamless promotion from Dev → Test → Prod.
They also detailed their KMS Ingestion Pipeline:
-
Trigger: S3 upload triggers Lambda/Step Functions.
-
Workflow: Runs Typhoon OCR to extract text.
-
Processing: Splits text into chunks.
-
Embedding: Uses Bedrock to generate embeddings.
-
Storage: Saves embeddings into OpenSearch vector DB for semantic search.
-
Reliability: Built-in retries and failure notifications.
HRM Chatbot workflow:
-
HR uploads files → OCR
-
System tags and organizes documents → Knowledge Base
-
Chatbot answers based on structured knowledge
Inteltion also plans to fine-tune models further and expand to new use cases.
5. NitiBench: Benchmarking Thai Legal AI
Chompakorn Chaksangchaichot, VISAI.AI
VISAI presented their research on NitiBench, a benchmark for evaluating LLMs in Thai legal tasks, alongside experiments using Reinforcement Learning (RL) to improve performance.
NitiBench powers “Seemumueang,” a legal search and analysis tool, and serves as a key dataset for Legal RAG.
Typhoon was included in the benchmark since it offered the largest Thai context window at the time of testing. Results showed Typhoon achieved comparable performance to leading models, especially strong on the NitiBench-Tax dataset.
Interested readers can download the full technical presentation here
Closing
Beyond the talks, the meetup also offered networking sessions for onsite attendees, and even an open-mic slot for anyone to share their open source projects—not limited to Typhoon. The goal: to foster knowledge exchange and empower Thailand’s open-source AI community.
We’re excited to host more events like this in the future—full of learning, sharing, and inspiration. 🚀
