Wow, that coding demo was super impressive. Can't wait to try it out. Just need to wait for the hardware market to cool down a bit so we can run these on our own commodity hardware.
My current feedback on Inkling 2 Small: 1. it feels intelligent, and can solve open problems throw at it like other quasi-frontier models (DeepSeek 4 Flash 0731 / DeepSeek 4.1 Flash 1017 / GLM 5.3 Flash etc). 2. its responses sometimes can be odd (they are honest about it in the blog post!), can have wrong use of markdown syntax, sometimes overly brief, etc. 3. tool uses are good, but sometimes have some old take on tools (too quickly to assume image generation tools cannot do text etc).
EDIT: by "open problems", I meant the problems I don't really formulate how to solve it first in my mind and then throw to it. I just describe what I want and let it to figure it out.
In terms of the active parameters and token/s on slow or less capable hardware it seems like it should be in the same general performance category as qwen 3.6 35B A3B.
From a research perspective, I'm even more interested in the varlooped variant they're releasing https://thinkingmachines.ai/model-card/tml-d24-varloop-resea... I'm not aware of any other working open source models like that.
EDIT: by "open problems", I meant the problems I don't really formulate how to solve it first in my mind and then throw to it. I just describe what I want and let it to figure it out.
https://huggingface.co/thinkingmachines/Inkling-2-Mini
https://huggingface.co/unsloth/Inkling-2-Mini-GGUF
In terms of the active parameters and token/s on slow or less capable hardware it seems like it should be in the same general performance category as qwen 3.6 35B A3B.