Inkling

Open weights 975B multimodal model built for fine-tuning

Artificial Intelligence

Development

Visit WebsiteSee on Product HuntHugging Face ⧉Twitter ⧉

the framing of being explicitly not the strongest model but the best base to adapt is refreshing, most labs oversell the raw benchmark score. with 41B active params out of 975B total, what's the realistic hardware floor for someone fine-tuning this through Tinker, is this something a well-funded startup can do on rented compute or does it really need lab-scale infrastructure?

LoRA is great, but supporting full fine-tuning or QLoRA as a built-in option would make this way more useful for folks working on smaller models where LoRA just doesn't cut it.

finally an option that lets me skip the gpu setup dance and just get to the actual tuning. ran a small lora job last night and it just worked, which honestly surprised me

one thing that would make tinker a lot more useful for me is built in support for evaluating models right after fine tuning, like running a small benchmark suite automatically so you can see if your lora actually helped without wiring up a separate eval pipeline

Really interesting approach. What's the minimum dataset size you'd recommend for effective domain fine-tuning?

As someone who's never trained an AI before, it seems pretty approachable. You fine-tune open models on your own data, and they handle the heavy infrastructure stuff. Inkling also feels like it's meant to be adapted to your workflow. Good idea

Built a quick LoRA job on Tinker yesterday and the setup was honestly painless. One thing that would be a huge help though: a built-in diff viewer or summary that shows what changed in the merged adapter weights so I can sanity-check before pushing to prod without having to script it myself.

Would love to see built-in support for evaluating checkpoints mid-training, so we can compare LoRA adapters on a validation set without writing custom eval loops. A simple callback or webhook when a checkpoint saves would go a long way for experiment tracking.

Love that Thinking Machines put the honest framing up front, "this isn't the strongest model today, it's a base you shape around your own work." Most open-weights drops oversell the benchmark line, so leading with adaptability instead is a cleaner pitch, and Apache 2.0 on a 1M-context multimodal MoE is a real gift.

The thing I keep bumping on is that "open weights" and "actually touchable" aren't the same at 975B. Realistically, who fine-tunes a model this size outside of Tinker? Wondering whether the openness is meaningful in practice, or whether downloadable weights are mostly a trust signal and Tinker is the real on-ramp most people will have to take to do anything with it.

Most of our pain when we LoRA an open model comes from the post-training rather than the weights. You tune for a narrow extraction job and the model keeps sliding back into the chatty assistant phrasing it picked up in RLHF, and you burn epochs beating that out. Is there a raw pre-trained checkpoint alongside the released one, or is the Apache 2.0 drop the post-trained model only? For something billed as a base to build on, that distinction decides a lot.