SCBX Group utilizes Gemma 3 12B to improve the performance, size, cost-efficiency, and Thai language proficiency of its Typhoon model Typhoon 2.1 brings improved Thai understanding in a smaller and more performant model with the help of Gemma SCB 10X, the technology investment arm of SCBX Group, is accelerating AI innovation across Thailand and the Association of Southeast Asian Nations (ASEAN) through its Typhoon initiative.Recognizing the urgent need for powerful, cost-efficient, and accessible Thai-proficient Large Language Models (LLMs), SCB 10X fine-tuned Gemma 3 12B to create Typhoon 2.1 Gemma.This model significantly advances the Typhoon family—which includes LLMs, VLMs, and speech recognition models—by offering cutting-edge performance to developers, startups, and research communities throughout the region. The challengeCreating accessible AI for millions users requires foundational models that are compact, efficient, and offer strong multilingual capabilities. The SCB 10X team launched the Typhoon initiative after identifying a gap in the market: existing solutions were often too large, prohibitively expensive to deploy, or lacked a deeper understanding of Thai linguistics.While previous Typhoon models were effective, the team sought to improve both Thai proficiency and cost-per-inference. Their prior flagship model, the Llama 3.1-based Typhoon 2 70B, incurred a high inference cost of $0.88 per million tokens, limiting its scalability. “ Gemma 3 provided the best balance of compact size and strong base performance, making it highly suitable for local deployment and scalable production use cases. Kunat Pipatanakul, Lead AI Scientist at SCB 10X The solutionWhen planning the next iteration of Typhoon, the team evaluated several open models—including Qwen3, Llama 4, and Gemma 3—to serve as the base. They found that models in the Llama 4 family were impractically large, while Qwen3, though effective, tended to "overthink" tasks, making it less ideal for general-purpose applications. Gemma 3 12B emerged as the clear choice, offering the best size-to-performance ratio, lower inference costs, and a high baseline understanding of Thai and other Southeast Asian languages.To transform the base model into Typhoon 2.1 Gemma, the team employed a multi-stage process:Supervised Fine-Tuning (SFT): The model was first trained on a curated dataset. Using a high-quality text classifier, the team filtered for culturally relevant Thai text and specialized content from higher education to capture deep linguistic nuance.Model Merging: The team then used a model merging technique to integrate the instruction-following capabilities of their previous Typhoon 2 model. This step ensured alignment with established Thai user preferences and behavior.Reinforcement Learning (RL): Finally, RL was applied to correct any artifacts from the merging process and refine the model’s ability to perform controllable, long-form reasoning.To maximize accessibility, the team hosts it as a managed api via Together AI. For developers who require more flexibility or local deployment, Typhoon 2.1 Gemma is available via Ollama, vLLM, or Llama.cpp. “ What sets Typhoon 2.1 Gemma apart is its balance of strong performance and lightweight cost-efficiency, making deployment practical in both enterprise and consumer environments. Kunat Pipatanakul, Lead AI Scientist at SCB 10X The impactAccording to Kunat, the in-depth Thai training and flexible deployment options "produced a model that is both highly accurate in Thai tasks and lightweight enough for efficient deployment, making it practical for real-world business applications."The results demonstrate a massive leap in efficiency. Typhoon 2.1 Gemma (12B) outperforms the previous Typhoon 2 (70B) in Thai, achieving a higher MT-Bench score with a model that is nearly 6x smaller.This improved performance comes at a fraction of the cost. Typhoon 2.1 costs just $0.20 per million tokens—a 77% cost reduction compared to Typhoon 2. Typhoon 2.1 Gemma achieves strong Thai and multilingual proficiency on MT-Bench benchmarks. Within four months of its launch, Typhoon 2.1 achieved 3.8 million downloads and is averaging 1 million API calls per month—a 70% increase over Typhoon 2’s monthly average.InnovestX is already using the model for customer service automations, while the Thailand Development Research Institute is using it to aid in document processing and scalable content analysis. Typhoon 2.1 is even being used in the education sector to power an exam prep assistant in partnership with the Office of the Education Council. What’s next SCB 10X plans to continue building with Gemma because of its robust toolset and wide-ranging compatibility. "We’ve benefited from a strong ecosystem of support and abundant resources, like open-source frameworks, that have accelerated our development process," added Kunat. The team is currently exploring enhancements to their models' agentic capabilities, expanding into multimodal AI with audio, and developing high-quality embedding models for Thai and multilingual use cases. These embeddings will open up more powerful applications in semantic search, recommendation, and knowledge retrieval, creating new opportunities for businesses and individual developers. Kunat emphasized that Gemma will likely be at the core of these new projects, stating, "We see strong ongoing potential in Gemma as a foundation for future developments."