Models

Full-song music generation with lyrics in 17 languages; about four minutes of audio in twenty seconds on an A100. Apache-licensed weights, unlike most music models.

Tencent's lightweight 8.3B video model with strong motion quality on a single GPU. Community licence permits commercial use under 100M monthly users outside the EU, UK and South Korea.

22B video model that generates synchronized audio in the same pass, up to 4K. Weights are open; commercial use is free below US$10M annual revenue, which ThaiRouter is.

Alibaba's mixture-of-experts video model, the most permissively licensed of the top open video models. 720p text-to-video and image-to-video; Wan 2.5+ are API-only and not open.

17B open image model with strong prompt adherence and photoreal output. MIT-licensed weights; Dev and Fast distillations exist for cheaper tiers.

20B image foundation model known for complex text rendering inside images. Qwen-Image-Edit shares the licence and will follow for image-to-image editing.

6B single-stream DiT from Alibaba's Tongyi-MAI team. Debuted #1 open-weights on the Artificial Analysis image arena; renders bilingual text and generates in under a second on datacenter GPUs.

Black Forest Labs' small FLUX.2 for sub-second generation and editing. The 4B checkpoint is the Apache-licensed one; the 9B is non-commercial and not offered.

8B text-to-speech from Fudan's OpenMOSS lab covering 31 languages including Thai, with zero-shot voice cloning and 48 kHz stereo output.

Speech recognition for 52 languages and dialects including Thai, with language detection and word timestamps. Ships with a vLLM-based serving stack.

Whisper large-v3 fine-tuned on Thai speech by SCB 10X's Typhoon team. Inherits Whisper's MIT licence.

OpenAI's Apache-2.0 open-weights MoE (117B total, 5.1B active). Configurable reasoning effort, native tool use, runs on a single 80 GB GPU.

Meta's open-weights Muse family model in its high-effort reasoning mode. Balanced quality and cost for general assistants.

NVIDIA's largest open Nemotron model, trained for enterprise agents with strong tool calling and long-form generation.

Compact open-weights generalist tuned for fast, low-cost chat and summarisation workloads.

Dense 27B model from the Qwen3.8 family with extended thinking. Best quality-per-parameter in its class; ideal for chat, RAG and tool use.

DeepSeek's flagship open-weights reasoning model. Top-tier on coding, math and agentic benchmarks at a fraction of proprietary prices.