Thinking Machines publishes Inkling 2 Small weights a week after launch
Thinking Machines Lab published the weights for Inkling 2 Small on Hugging Face on November 12, ending the model's first week as a hosted-only release. The multimodal mixture-of-experts model has 594 billion total parameters with 17 billion active per token, up to 1 million tokens of context, and scores 31 on the Artificial Analysis Intelligence Index. The weights arrive under Apache 2.0 in NVFP4 and MXFP8 quantisations, a week after the November 5 launch through the lab's Tinker platform.
📊 Scorecard: our Bullish call on $NVDA is scored from $166.19, the last close before it ran. Its first result comes after the next close.
The News
Thinking Machines Lab published the weights for Inkling 2 Small on Hugging Face on November 12, ending the model's first week as a hosted-only release. The lab introduced Inkling 2 Small on November 5 through its Tinker platform, promising the weights under Apache 2.0 after outside safety testing. The November 12 upload delivers on that promise, in NVFP4 and MXFP8 quantisations — two compressed formats that let the model run efficiently on different hardware.
Inkling 2 Small is a multimodal mixture-of-experts model, which takes text and images in and produces text out. It has 594 billion total parameters with 17 billion active per token, up to 1 million tokens of context — a token is roughly a word or word fragment, so a million tokens means the model can take in very long documents in one go — and it scores 31 on the Artificial Analysis Intelligence Index. The model was trained on NVIDIA GB300 NVL72 systems.
The November 12 update also widens access. API access now runs through Together, NVIDIA, Fireworks, Baseten, Nebius, Modal and Cloudflare, and serving and training integrations cover Unsloth, SGLang, vLLM and TokenSpeed, plus post-training via Miles, SkyRL and Snowflake Cortex Training. One thing is missing: the lab has not published the results of the outside safety testing it cited as the reason for the week's delay.
Inkling 2 Small went from hosted-only to open weights in a week — the promise kept, the safety results still unseen.
Why It Matters
Open weights change who can use a model. When a lab only offers a model through its own hosted platform, developers rent access and play by the lab's rules. When the weights are published, anyone can download the model, run it on their own machines, fine-tune it for a specific job, and build products on top of it without asking permission. That is why the November 12 upload matters more than the November 5 launch: the launch announced the model, but the weights release hands it over.
The size numbers tell the same story in a different way. A mixture-of-experts design means the model holds 594 billion parameters in total but only switches on 17 billion of them for each token it processes. In plain terms, it is a very large model that behaves like a much smaller one when it runs, which keeps costs down. A million tokens of context means it can swallow long documents — whole codebases, long reports — in a single pass, and the 31 score on the Artificial Analysis Intelligence Index is the lab's reported score on an independent benchmark.
The distribution is unusually wide for a week-old model. Seven API providers and a full set of serving and training tools on day one means a developer can pick almost any familiar stack and start using Inkling 2 Small immediately, rather than waiting for integrations to appear. That breadth is a deliberate choice by the lab, and it is the difference between an open release that gets used and one that sits on a download page.
The gap in the story is the safety testing. The lab said the week's delay was for outside safety testing, and it has not published the results. That does not mean the testing failed — but for a model this large, released openly, the results are the evidence that the delay bought something. Until they appear, the release is a promise kept on timing and an open question on transparency. Thinking Machines is private, so the listed read-through is NVIDIA, whose GB300 NVL72 systems trained the model and whose platform serves it.
Analyst's Take
Our call is Bullish. The November 12 weights publication is the event, not the November 5 launch: a frontier-scale open model that anyone can download, run and build on is a real step for the open-weights ecosystem, and the breadth of the rollout — seven API providers and a full serving and training stack on day one — shows the lab planned for adoption, not just a headline. For NVIDIA, whose GB300 systems trained the model and whose platform now serves it, every open-weights release that runs on its hardware is a demand tailwind.
What to watch: whether the lab publishes the outside safety-testing results it cited for the week's delay, and how quickly developers pick the model up across the new providers. The 31 is the model's score on Artificial Analysis's independent Intelligence Index, as reported by the lab, so independent testing will matter more than the launch-day number. If adoption follows the access, this becomes a durable open option; if not, it is a well-distributed release that few use.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What does it mean to publish the weights?
The weights are the trained numbers inside a model — what it learned during training. Publishing them lets anyone download the model and run it on their own hardware or service, instead of only using it through the lab's hosted platform. The weights are released under Apache 2.0, an open licence that permits commercial use.
What is a mixture-of-experts model?
It is a model built from many smaller specialist sub-models, called experts. For each token — roughly a word or word fragment — only a few experts switch on, so the model can be huge in total size while only a small part runs at any moment. Inkling 2 Small has 594 billion total parameters but only 17 billion active per token.
What are NVFP4 and MXFP8 quantisations?
Quantisation shrinks a model's weights into smaller number formats so it runs faster and uses less memory. NVFP4 and MXFP8 are two such formats offered for this release, letting different hardware setups run the model efficiently.
Why did the weights take a week?
The lab introduced Inkling 2 Small on November 5 through its Tinker platform and promised the weights under Apache 2.0 after outside safety testing. The November 12 upload delivers on that promise — but the lab has not published the results of the outside safety testing it cited as the reason for the week's delay.
How can developers use Inkling 2 Small now?
The November 12 update widens API access through Together, NVIDIA, Fireworks, Baseten, Nebius, Modal and Cloudflare, and adds serving and training integrations including Unsloth, SGLang, vLLM and TokenSpeed, plus post-training through Miles, SkyRL and Snowflake Cortex Training.
👁
Source: Superpower Daily