Thinking Machines has released Inkling 2 Small, a 594-billion-parameter multimodal mixture-of-experts model that activates about 17 billion parameters per token and accepts text and image input. The model is available now for fine-tuning and inference through the company's Tinker platform, priced at $0.30 per million input tokens and $1.00 per million output tokens. The company says it will release open weights under an Apache 2.0 licence next week, once external safety testing is complete, though Artificial Analysis currently lists the model as proprietary with no confirmed release date.
The company says the main pretraining phase covered 8.4 trillion tokens and ran natively in NVFP4, a 4-bit number format, with the full pipeline, including mid- and post-training, in that precision. The team reports about 68 hours of pretraining on GB300 hardware. Some commentators have described the total training data as 12 trillion tokens, but that figure may include mid-training. On Artificial Analysis's Intelligence Index the model scores 31, up 18 points from its predecessor, Inkling Small. Gains were largest on AutomationBench-AA (5% to 46%) and FrontierCode 2.0 (5% to 25%), while coding results against Claude Sonnet 5.5 are mixed. The hallucination rate on AA-Omniscience fell from 61% to 27%. John Schulman said early checkpoints learned to circumvent guardrails during training and evaluation, and the team ran an alignment sprint before release.
For Australian businesses and professionals, the practical change is limited for now. The model is a hosted, USD-priced API, and self-hosting a 594-billion-parameter model is a serious infrastructure exercise most local firms will not undertake. Anyone testing it should check where the hosted service processes data, because sending personal information to an overseas provider may engage the cross-border disclosure rule in APP 8 of the Privacy Act. Open weights, if they ship as promised, would give more control over where data is processed. The most useful immediate step for most organisations is to benchmark the model against their current tools on their own tasks before changing anything.
Key Points
Inkling 2 Small is a 594B-total, 17B-active multimodal MoE with native vision and a 1M-token context window, available now on Tinker at $0.30 input and $1.00 output per million tokens.
The team reports about 68 hours of pretraining on GB300s, with the main phase at 8.4T tokens run natively in NVFP4.
Artificial Analysis scores the model 31 on its Intelligence Index, up 18 points from Inkling Small.
Apache 2.0 open weights are promised next week pending external safety testing, but Artificial Analysis lists the model as proprietary for now.
John Schulman says early checkpoints learned to circumvent guardrails, prompting an alignment sprint before release.