Thinking Machines Lab’s newly open-weighted Inkling 2 Small model took the No. 1 spot among US open models on both Agent Arena and Code Arena WebDev while rolling out on Cloudflare Workers AI, vLLM, TokenSpeed and SkyRL. In Agent Arena, the 594B parameter model ranked second among all open models globally across 11,500 real-world sessions with the highest open-model steerability at +10.9% and a median cost of $0.26 per task, compared with $1.22 for top-ranked Kimi K3.1 Pro. It also scored 1,647 points to place fourth among open models globally in Code Arena WebDev and ranked first among open-weight models on DesignArena’s Overall HTML leaderboard with an Elo of 1,384, 32 points ahead of Kimi K3.

On the inference side, vLLM and Inferact added support reaching up to 630 tokens per second per user at TP4, using mixed NVFP4 and MXFP8 expert layers across the model’s 17B active parameters and a DSpark speculative decoder to fit the 394 GB checkpoint on a single Nvidia GB200 or GB300 node. TokenSpeed posted 476 tokens per second per user on four Nvidia B300 GPUs and added AMD Instinct MI350X support, while SkyRL validated reinforcement learning with vLLM and Megatron. Cloudflare noted that Thinking Machines used Inkling 2 Small on its Tinker platform to autonomously train Qwen 3.5-4B for 4.5 hours, more than doubling its forecasting score.