Thai Open AI Project

Model Hub

AI models

Every model card states its base, task, hardware requirement, benchmark status and known limitations.

drone-detect-th

EXPERIMENT

Detection model trained on Drone Dataset TH for small, distant aerial targets under tropical sky conditions.

BASE
YOLO family
PARAMS
~11M
TASK
Object Detection
LANGUAGE
—
QUANTIZATION
FP16 / INT8
HARDWARE
Single GPU or edge accelerator
LICENSE
Research use
VERSION
v0.4

Benchmark

  • Evaluationper-weather split
  • Target latencyreal-time edge

Limitations: Not validated for airspace-safety-critical decisions

thai-local-8b-q4

EXPERIMENT

Quantised local deployment configuration studied for Thai-language quality per watt on commodity hardware.

BASE
Open-weight 8B LLM
PARAMS
8B
TASK
Text generation
LANGUAGE
TH / EN
QUANTIZATION
Q4_K_M
HARDWARE
Consumer GPU, 8GB VRAM class
LICENSE
Depends on base model
VERSION
v0.2

Benchmark

  • Thai language scoreunder evaluation
  • Tokens/sechardware dependent

Limitations: Quantisation degrades long-form Thai reasoning; quantified in ongoing runs

thai-open-router-1

PROTOTYPE

Routing policy that selects between local and cloud models per task class, used by the UACP-HMA reference stack.

BASE
Policy router (non-generative)
PARAMS
—
TASK
Model routing
LANGUAGE
TH / EN
QUANTIZATION
—
HARDWARE
CPU
LICENSE
Open research
VERSION
v0.6

Benchmark

  • Routing classestask-class based

Limitations: Routing policy is tuned to the current model pool only

muaythai-action-v0

EXPERIMENT

Experimental strike and clinch classifier conditioned on estimated fighter pose sequences.

BASE
Pose-conditioned temporal model
PARAMS
~24M
TASK
Action Recognition
LANGUAGE
—
QUANTIZATION
FP16
HARDWARE
Single GPU
LICENSE
Research use
VERSION
v0.1

Benchmark

  • Strike classesin development

Limitations: Severe degradation under heavy clinch occlusion