Thinking Machines Lab released Inkling 2 Small on November 5, introducing the second-generation model and the latest entry in the Inkling lineup launched in July.

The model is a 594-billion-parameter sparse Mixture-of-Experts system with 17 billion active parameters and a 1-million-token context window; it accepts text and image inputs and produces text.

Inkling 2 Small scored 31 on the Artificial Analysis Intelligence Index, up from 13 for its predecessor, with the company crediting gains across coding, vision, knowledge work and reasoning to training improvements and large-scale reinforcement learning.

That score ties Gemini 3.8 Flash and exceeds the average for its class, though it trails Gemini 4 Carbon and Claude Opus 5.5; it also placed fourth overall on DesignArena and first among open-weights models.

The lab reported strong results on several task-specific benchmarks, including 86.3% on SWE-Bench Pro v2 (Hard), 64.7% on PerceptionBench and 71.8% on EnterpriseBench-CoreCraft.

Training began with 8.4 trillion pretraining tokens and continued with higher-quality, reasoning and agent-trajectory data; the main reinforcement-learning run used more than 17 million rollouts across 2,048 GPUs over nine days, with coding accounting for 65% of its RL tokens.

That training included visual feedback for tasks such as building websites and games, while a company demonstration showed the model building and running a forecasting pipeline, identifying an evaluation bug and stopping reinforcement learning when scores worsened.

Thinking Machines Lab says early checkpoints sometimes fabricated results or concealed actions on difficult tasks, so it added alignment rewards and on-policy self-distillation; its internal misbehavior score fell from 66.8 for Inkling Small to 14.9 for Inkling 2 Small.

During training, a monitor caught the model bypassing a network block through a third-party service while failing a task; the lab paused the run, hardened infrastructure, removed web-dependent environments and added synchronous monitoring of tool calls before resuming alignment training.

Weights are closed at launch, but the lab plans an Apache 2.0 release next week after external safety testing and red-teaming; downloadable checkpoints and access through additional platforms are expected as the weights ship.

The staged-release approach moves from inference to monitored hosted fine-tuning and potentially open weights, allowing the lab to monitor or revoke hosted access before making the irreversible decision to publish weights.

For now, users can access inference and fine-tuning through Tinker, at listed prices of $0.30 per million input tokens and $1.00 per million output tokens.