Inkling 2 Small trained in 68 hours, but its open weights still await safety tests

Thinking Machines Lab has introduced Inkling 2 Small, a multimodal Mixture-of-Experts model with 594 billion total parameters and 17 billion active at runtime. The company says pretraining from scratch took 68 hours on fewer than 12 trillion tokens using native NVFP4.

A 594 billion parameter model activates only 17 billion per token

The MoE architecture does not activate all 594 billion parameters at once. This reduces inference compute while retaining a large total capacity. Inkling 2 Small accepts image input through a 292 million parameter vision encoder and supports variable reasoning effort.

It scored 31 on the Artificial Analysis Intelligence Index. Thinking Machines has made the model available for fine-tuning and inference through Tinker. At launch, however, it remains closed weight. Full weights are planned for the coming weeks, pending additional safety tests.

Fast training shifts the contest toward process efficiency

The result matters less as a race for the largest parameter count and more as a demonstration of low precision training, MoE routing and a joint RL run. Fewer than 3 days of pretraining do not represent the full cost of developing the model. Data preparation, experiments, post-training and safety evaluation sit outside that number.

For teams that want to adapt a model, the combination of 17 billion active parameters, multimodality and planned open weights is attractive. Practical value will depend on hardware needs, throughput and fine-tuning quality outside Tinker infrastructure.

The benchmark table shows strengths and conspicuous gaps

The model reached 90.2% on BrowseComp with context management and 64.7% on PerceptionBench. It trails elsewhere: Terminal Bench v4.0 shows 13.1%, compared with 33.8% for DeepSeek V4.1 Flash. The results do not support a simple claim that Inkling 2 Small beats comparable models across the board.

Most safety evidence comes from the developer and selected external evaluations. The company reports large gains over earlier Inkling models, yet it is keeping the weights closed until more testing is complete. That is a sensible brake and a reminder that announcing an open model is not the same as releasing open weights.

The real test begins when the weights leave Tinker

Reproducible benchmarks, hardware requirements and fine-tuning in independent environments will decide the result. It will also matter whether the full-weight release date slips and what use the eventual license actually permits.

If outside teams confirm the performance at 17 billion active parameters, the 68-hour training run will be a strong signal for more efficient scaling. For now, this is an impressive developer report whose most important artifact is still missing.

Lilith's verdict

The stopwatch reads 68 hours, but the crate of weights remains locked in the lab. The real race starts when outside teams get the crate and measure the full bill.