Thinking Machines releases open-weights model Inkling, a 975B MoE with self-fine-tuning demo on Tinker
- Thinking Machines released Inkling, an open-weights Mixture-of-Experts transformer with 975B total parameters and 41B active, pretrained on 45 trillion tokens of text, images, audio and video, supporting up to a 1M token context window.
- A smaller preview model, Inkling-Small, has 12B active parameters (276B total per HN discussion) and is still being tested before its open-weight release.
- The model is available for fine-tuning on Tinker, and Thinking Machines demoed Inkling writing its own fine-tuning job (a lipogram task banning the letter 'e'), running it, evaluating the result, and swapping in the new checkpoint itself.
- HN commenters note Inkling underperforms GLM 5.2 in agentic and coding benchmarks despite being about 30% larger, though it multimodal support and long context draw interest as a possible DeepSeek V4 Flash style replacement once Inkling-Small ships.
- Weights are on Hugging Face listed as Apache-2.0, but commenters flag that Thinking Machines also applies an AUP (acceptable use policy) alongside the license, raising licensing clarity questions.
Hacker News opinions
America needs its own DeepSeek or Z.ai, and a lot of us root for open Chinese models to win because we have no other choice. Thinking Machines might actually be it.
It's not as good as GLM 5.2 for agentic workflows while being bigger. Switching costs are basically zero so competition here is going to be ruthless.
There's a bunch of companies chasing open weights already, Arcee, Reflection, Llama. Not sure the fine-tuning-plus-open-weights combo is a real moat if it's just LoRA under the hood, but I'll watch it.
What's the actual business model for an open weight model anyway?
I never thought I'd see the day they released an actual model instead of just another blog post. Jokes about the localhost demo aside, this is the best western open weights model I've seen.
If it's ~30% bigger than GLM 5.2 and still worse, why would I bother fine-tuning this one? Maybe the multimodal angle is the draw.
Benchmarks never tell the whole story, some open models are just benchmaxxed. Having real vision input is genuinely useful for a lot of tasks regardless of the score gap.
There's an Inkling-Small at 276B total, 12B active, way smaller than GLM 5.2 and still multimodal. Not released yet but could be a good DeepSeek V4 Flash replacement once it drops.
This seems particularly strong at instruction following but weaker at coding than others. Still nice to get more diversity in open weights, I'll test the personality myself.
Long context plus multimodal audio is nice, but MiniMax M3 and DeepSeek v4-Pro already do long-context multimodal well, and honestly all long-context claims are a trap since performance falls off hard past 150k to 200k tokens.
That 276B A12B version might fit in 128GB once quantized to 2 bits, which makes it pretty interesting if they ever release the weights.
For most part it's better than Nemotron, worse than GLM, so this might be the best American open weights model right now.
It's almost double Nemotron 3 Ultra's size so it should be better, though its active params are actually lower at 41B versus 55B.
Too bad we'll never really know how good it is since they only showed a radar plot for benchmarks.
Serving the model over API is the obvious play, plenty of open weight companies already do that as their main revenue source.
Thinking Machines might be the only major US lab whose business interest actually lines up with open sourcing strong customizable models, since their fine-tuning API needs a good base model to sell against.