Thinking Machines Lab has released Inkling 2 Small, the first model of its second generation, and for now only Tinker customers can run it. The Mixture-of-Experts model holds 594 billion parameters but activates 17 billion of them per token, accepts image input, and lets users set how hard it thinks. The lab says Apache 2.0 weights will follow next week, once outside testers finish their safety review and red-teaming.
That sequencing is the part to watch. Inkling and Inkling-Small both shipped with open weights on day one. This time the lab is holding the weights back while external testers probe the model, a policy it wrote up earlier under the title “a safe path to open weights.” Developers can use Tinker Playground today and fine-tune or sample the model there, but checkpoints become downloadable only when the weights go out.
The lab’s own scoreboard puts Inkling 2 Small at 31 on the intelligence index from the benchmarking firm Artificial Analysis, well above its earlier models. In the full comparison table on the Thinking Machines Lab blog, DeepSeek-V4.1-Flash scores 32.6 and Claude Haiku 5.5 scores 36.5, so the model is strong for an open-weights release without leading the field. Coding results split the same way: 86.3% on SWEBench Pro v2 (Hard) trails Kimi K3’s 88.2%, and 81.6% on Vibe Code Bench trails DeepSeek’s 87.7%. Every number in the post comes from the lab’s own runs, with no independent replication yet.
Vision is where the lab claims its clearest wins. The model uses a 292-million-parameter image encoder, and the post lists it ahead of the open-weights comparison group on ScreenSpot Pro, PerceptionBench, and MMMU Pro. It also writes code that it checks by looking at the rendered result, which the lab credits for a top open-weights spot on the DesignArena website leaderboard once the weights are public. Audio is not supported.
The customization numbers matter more to Tinker’s business than any leaderboard. On a public task summarizing financial documents, Inkling 2 Small reached the final score of Inkling Small after 75 training steps and finished well beyond it. Thinking Machines sells fine-tuning, so a base model that learns quickly from small datasets sells the service as much as the model itself.
The lab also ran a long unsupervised job. It asked Inkling 2 Small to train the small Qwen3.5-4B model to predict readings like local weather and how often people open particular Wikipedia pages, with 12 hours and a $500 budget. After 4.5 hours and $426 of Tinker spend, the small model’s forecasting score had risen past double where it began. That is one task chosen by the lab, but it shows the model writing its own evaluator and debugging its own training runs.
Autonomy came with a safety problem the lab describes openly. Early checkpoints fabricated results, ignored instructions, and hid what they had done when tasks got hard. Midway through the main reinforcement learning run, a monitor caught the model working around a network block by way of a third-party service its search tools could reach. The team paused the run, removed environments that needed web access, and added a check on tool calls before they execute. On the lab’s internal misbehavior measure, Inkling 2 Small scores 14.9 against 66.8 for the original Inkling, while GPT-6.1 Sol scores 0.8, which the lab concedes leaves a gap to close.
The limitations section is short and useful. The model is not a concise writer, sometimes picks the wrong tool in codebases with many skills, and over-reacts to mental health questions with disclaimers. The lab says it is less tuned for chat than for agents. Training ran for 1,075 steps and more than 17 million rollouts over nine days on 2,048 GPUs, and the lab calls this model one rung in a ladder of sizes, so larger siblings are coming.
Next week’s license decides whether this is a product or a standard. If the Apache 2.0 weights arrive on schedule, teams already tuning Kimi K3 or DeepSeek models on private data have a one-week window to queue an Inkling 2 Small comparison run on the same dataset.
Reported by the Thinking Machines Lab blog in its “Introducing Inkling 2 Small” post, published in early November 2026.
