Compare Capabilities กำลังพัฒนา

Pick the Right Extension for Your Task

Nara AI has a single base model (12B) — you choose pluggable LoRA extensions for your use case

Base Model

One Model, Runs Everywhere

Nara AI 12B กำลังพัฒนา

Built on Gemma 4 12B Context 32K Apache 2.0 GPU 16GB (Q4) or Mac (MLX)

Weights not yet publicly released — training and tuning in progress

LoRA Extensions

Plug and play — no need to retrain the whole model

LoRA	Capability	Status	Best For
`cpt-th`	Thai Language Foundation CPT on ~144M token Thai corpus — read and write Thai	กำลังพัฒนา	General Thai-language tasks
`sft`	Instruction Following SFT for instruction-following — conversation, summarization, translation	กำลังพัฒนา	Chatbot, assistant, RAG
`agent`	Tool Use Trained to call tools and follow multi-step procedures (agent/tool-use)	แผนงาน	Workflow automation
`law`	Legal Domain Specialized extension for Thai legal contexts	แผนงาน	Legal work, contracts

Real Results

Development-stage scores — not a comparison against other closed models

Benchmark	Nara AI	Notes
Belebele Thai (Reading Comprehension)	0.635	Higher than Typhoon (0.390)
MMLU (General Knowledge)	0.680
CEval (Chinese Evaluation)	0.762

We only release scores we can actually measure — we do not fabricate comparisons against models we have not tested ourselves

Guidance

Tell us what you want to do, we will recommend the right extension

Use the sft LoRA on the 12B base model — ideal for Q&A tasks

→ base + sft

Use the cpt-th LoRA for deeper Thai language understanding

→ base + cpt-th

Add the law LoRA for legal terminology and context

→ base + cpt-th + law

Add the agent LoRA to call APIs and follow procedures

→ base + agent

Start with the base model + cpt-th LoRA — it is the best starting point for Thai language tasks