Thinking Machines Lab released its second model generation on Thursday, and the benchmark gap to its own first effort is the headline.
Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index. Inkling Small, released in August, scored 13. The new model packs 594B total parameters, keeps only 17B active per request, and accepts images alongside text.
For now it runs only through Tinker, the company’s fine-tuning service, with weights held closed. Thinking Machines intends to publish them under Apache 2.0 next week once outside red-teaming finishes.
Built for long agent runs
The training recipe leaned on reinforcement learning across hundreds of thousands of long coding tasks, spread over several agent harnesses and context-compaction strategies. Graders judged how a model approached a problem, not just whether its patch passed tests.
Coding scores land near the top without taking first place. On SWEBench Pro v2 Hard it reaches 86.3%, behind Kimi K3 at 88.2%. FrontierCode v2.0 comes in at 25.4% against DeepSeek-V4.1-Flash-1017’s 28.5%.
The gains concentrate in autonomy. The lab showed it building a full city simulator in Three.js with no pre-made assets, generating buildings, textures and traffic procedurally while tracking power and budgets. Early checkpoints were tested on OpenCode under the name Fledge-Alpha.
Where it still falls short
Thinking Machines is candid about the limits. The model is not a concise writer, sometimes picks the wrong tool in codebases stuffed with skills, and is noticeably weaker as a chat partner, overusing slashes and overreacting to safety-flavored prompts. Audio input is absent entirely.
The staged release follows the lab’s published policy on opening weights safely. If the safety pass holds, Inkling 2 Small becomes one of the largest openly licensed models available, and a direct challenge to the Chinese open-weight leaders it still trails on several coding boards.
