Inkling-Small Is a Quarter the Size of Its Predecessor and Nearly as Capable
Thinking Machines has released a second open-source AI model just two weeks after its first. The smaller version costs less to run, scores higher on several coding tests, and comes with a business-friendly licence.
Key points
- Thinking Machines released Inkling-Small on 9 July 2025, just two weeks after launching its original Inkling model.
- Inkling-Small has 276 billion total parameters but uses only 12 billion at a time, compared with 975 billion total and 41 billion active for the original Inkling.
- Third-party benchmarking site Artificial Analysis scored Inkling-Small 40 on its Intelligence Index, one point below Inkling's 41.
- Inkling-Small outscores Inkling on coding tests, including 80.2% versus 77.6% on SWE-bench Verified, a standard coding benchmark.
- The model is released under the Apache 2.0 open-source licence, which lets companies use, modify and commercialise it with few restrictions.
Two weeks. That is all it took Thinking Machines, the startup founded by former OpenAI chief technology officer Mira Murati, to follow up its debut AI model with a smaller, cheaper, nearly-as-capable successor.
The new model is called Inkling-Small. It is a multimodal reasoning model, meaning it can take in text, images and audio and produce text in reply. Despite the name, it is not small by any ordinary measure, but it is dramatically leaner than its sibling, and the performance gap is surprisingly narrow.
How does it compare to the original Inkling?
On Artificial Analysis's Intelligence Index, a third-party scoring system that ranks AI models across dozens of tests, Inkling-Small scores 40 compared with Inkling's 41. One point of separation from a model that is roughly four times larger.
The size difference comes down to architecture. Inkling-Small uses a design called a sparse Mixture-of-Experts model. Think of it like a large team of specialists: the model has 256 specialist sub-networks, called experts, but for any given word or phrase it processes, it only consults six of them plus two general-purpose experts that are always active. That is why the model holds 276 billion parameters (learned values baked in during training) in total, yet activates only 12 billion at a time. Less work per step means lower computing costs.
On several coding and reasoning benchmarks, Inkling-Small actually beats its larger sibling. It scores 80.2% on SWE-bench Verified, a test of real-world software bug fixing, versus 77.6% for Inkling. It also edges ahead on a science coding benchmark called SciCode and on Humanity's Last Exam, a test of expert-level reasoning.
The advantage is not universal. Inkling retains a clear lead on factual knowledge tasks: 23.7% versus 15.5% on one banking-domain test. Enterprises using the model for high-stakes factual queries will still need to pair it with retrieval tools and human review.
Can you actually run it yourself?
Not on a laptop. The full-precision version of Inkling-Small needs at least 600 GB of combined GPU memory, the kind of specialised computing power found in server rooms, not home offices. Thinking Machines lists two supported hardware setups: four NVIDIA B300 chips or eight NVIDIA H200 chips.
A compressed version cuts that requirement to roughly 180 GB, and the company says that variant can run on a single B300 chip. Still, this is firmly enterprise territory.
The "Small" label is relative to Inkling, not to the broader world of lightweight models that run on consumer hardware. That said, the full model weights are freely available on Hugging Face, the main platform for sharing open AI models, and the open-source community will almost certainly produce even more compressed versions in the coming weeks.
What does it cost to use via the API?
Thinking Machines is currently offering a launch discount, bringing pricing to $0.58 per million input tokens and $1.44 per million output tokens for the standard 64,000-context version. A token is roughly three-quarters of a word, so a million tokens covers a large volume of text.
| Pricing tier | Cost per million tokens |
|---|---|
| Input (prefill) | $0.58 |
| Output (sampled) | $1.44 |
| Training | $1.73 |
| Cached input | $0.116 |
The Apache 2.0 licence covering the model is worth noting. It allows companies to use, modify, fine-tune and build commercial products on top of Inkling-Small without the revenue thresholds or branding conditions that appear in some other AI model licences. For legal and procurement teams, that is a meaningful simplification.
Common questions
Does "open source" mean anyone can use this freely?
Apache 2.0 is one of the most permissive standard licences in software. Companies can use, modify and sell products built on Inkling-Small, as long as they follow the licence's attribution rules and review any acceptable-use policy Thinking Machines publishes alongside it.
Is this model safe to trust for important decisions?
Inkling-Small's factual knowledge scores lag behind the larger Inkling, and its score on one factual index is negative, meaning it falls below baseline on broad knowledge coverage. Any organisation using it for medical, legal or financial decisions should build in retrieval from verified sources and human review of outputs.
