Thinking Machines Lab published the open weights for its 594B parameter Inkling 2 Small model on Hugging Face under an Apache 2.0 license following external safety testing, adding serving and fine-tuning support on Together AI and SGLang. SGLang enabled initial serving at 576 tokens per second per user at batch size 1 on Nvidia Vera Rubin chips and 123 tokens per second at batch size 192, alongside reinforcement learning on eight Blackwell GPUs and compatibility with Nvidia Hopper, AMD MI350X and MI355X, and Google TPU v7x hardware. OpenCode also made the multimodal mixture-of-experts model, which uses 17B active parameters and a context window of up to 1 million tokens, available for free in select regions for a limited time.