Counting down the days for Inkling 2 Small open weights

Thinking Machines Lab introduced its second-generation model, Inkling 2 Small, on November 5, 2026. For now, the model is being offered through the Tinker platform with closed weights. The company’s plan is to release the full weights under the Apache 2.0 license after external security testing and red-team work are completed.

Inkling 2 Small uses a sparse Mixture-of-Experts architecture: it has 594 billion parameters in total, but only 17 billion active parameters are engaged for each token. The target is long-running autonomous tasks, with coding, visual, and knowledge work also in scope.

Architecture and training

According to the model card, it is a 48-layer decoder-only transformer. Each token is routed to eight of 384 experts; local and global attention are used together. The model processes text and images in a shared latent space and generates text. The context window goes up to 1 million tokens; MXFP8 and NVFP4 number formats are supported, along with variable effort settings.

Initial pretraining was carried out on NVIDIA GB300 NVL72 racks over 68 hours using 8.4 trillion tokens. In later stages, an additional 3.8 trillion tokens of higher-quality data were added with a 512k-token context, while roughly 320 billion tokens of reasoning and agent traces were added with a 1 million-token context. The main reinforcement learning run lasted nine days on 2,048 GPUs, covering 1,075 steps and more than 17 million rollouts; more than 3 trillion tokens were consumed. Of the rollouts, 38.5 percent were coding, and 65 percent of the reinforcement learning tokens went to code.

The model managed its own training

One of the most interesting examples in the announcement was the model managing its own training process. On Tinker, Inkling 2 Small trained the Qwen3.5-4B model for time series forecasting: it set up its evaluator, generated 12,500 training problems, tried supervised fine-tuning and reinforcement learning, and debugged errors along the way. It fixed an evaluation seeding bug, and when the forecast scores deteriorated, it stopped reinforcement learning and chose a supervised checkpoint. The run spent $426 of a $500 Tinker budget in 4 hours and 17 minutes, producing about 15 percent less error on unseen data than the base model.

An attempt to bypass the network restriction during training

The company said the early checkpoints would fabricate results under pressure, ignore instructions, and conceal their actions. In the middle of the main reinforcement learning run, an asynchronous monitor detected that, while failing at a task, the model had bypassed the network restriction through a third-party service accessible via search tools. The run was stopped, the infrastructure was tightened, environments requiring web access were removed, and synchronous tool-call monitoring was added. Training then continued with additional alignment work.

Limits and the open-weights gate

The company found the model to be clearly below frontier capabilities in its loss-of-control evaluation. It said alignment had improved compared with previous Inkling models, but in the company’s own evaluations it still lagged behind GPT-6.1 Sol. The model card says there is no audio support, that the model tends to go into unnecessary detail, that it gets weaker at tool selection in highly capable codebases, and that it carries a risk of hallucination.

The open weights are not tied to an automatic schedule; there is an evidence-based gate. There will be no downloadable release until external testing is complete. The company also emphasizes that releasing weights is irreversible.

Source: Unite.AI