Thinking Machines' Inkling 2 Small hits 31 on AA Index, weights next week

Inkling 2 Small arrives as a 594B-total / 17B-active multimodal MoE, roughly twice the parameters of Inkling Small, and is positioned as a step up in capability for the Tinker ecosystem. The model was pretrained in 68 hours on GB300s, with the main pretraining phase running 8.4T tokens natively in NVFP4. Post-training involved 17M rollouts over 9 days, and an engineer highlighted long-horizon capabilities bootstrapped from near-zero pass rates.

Customizability is framed as a design target, with Soumith Chintala saying it is 'probably one of the only model programs in the industry with a launch blocking eval around customizability.' He adds that for a fixed data and dollar budget, it fine-tunes better than alternatives. The model is token-heavy at inference — about 150K output tokens per Index task, compared with roughly 37K for Inkling Small — though low per-token prices keep cost per task at $1.18 versus $0.20 for its predecessor.

Vals ranks it #7 among open-weight models globally, noting strength in finance and legal but considerable ground to make up on technical agentic tasks. The open-weights claims remain conditional: Apache 2.0 weights are promised next week, and Artificial Analysis lists the model as proprietary at launch with an unconfirmed weights date.