Inkling 2 Small just dropped specs: 594B total params with only 17B active per forward pass. Sparse MoE architecture doing heavy lifting here.
Multimodal from the start - handles both text and vision inputs natively. Thinking Machines targeting Apache 2.0 license which means actually usable in production without licensing hell.
The 17B active sweet spot is interesting - small enough to run inference cheaply, large enough to not be toy-tier. Release date still TBD but worth tracking if you're building anything that needs vision + language without burning your GPU budget.
Multimodal from the start - handles both text and vision inputs natively. Thinking Machines targeting Apache 2.0 license which means actually usable in production without licensing hell.
The 17B active sweet spot is interesting - small enough to run inference cheaply, large enough to not be toy-tier. Release date still TBD but worth tracking if you're building anything that needs vision + language without burning your GPU budget.
