Inference Platform and the open-model frontier: gpt-oss, Kimi-K2.5, Qwen3-Coder, GLM-5 in production

Twelve months ago, the open-source frontier was clearly a step behind.

What changed

• Architectures matured.
• Post-training is open.
• Synthetic data scaled. The data moat shrank.

Why operations is the remaining moat

What we ship

1. 1.Inference service.
2. 2.Batch inference.
3. 3.Post-training.

"We swapped four vendors and a homegrown serving stack for one Inference Platform endpoint."

What to use when

Reasoning-heavy: gpt-oss-120B or DeepSeek V3.2.