Mira Murati's lab shrank its own model to a quarter of the size and lost one point
Thinking Machines released Inkling Small, an open-weights reasoning model with 276 billion parameters that scores 40 where the 975-billion-parameter original scores 41.
Two weeks after shipping its first model, Thinking Machines has released a second one that is much smaller and almost as good. Inkling Small is an open-weights reasoning model with 276 billion total parameters and 12 billion active, released under the Apache 2.0 licence. On the Artificial Analysis Intelligence Index it scores 40. The original Inkling, at 975 billion parameters, scores 41.
A few of those terms deserve translating. Parameters are the adjustable numbers a model learns during training, and more of them usually means a bigger, more expensive model to run. Total versus active matters here because both models use a mixture-of-experts design, where only a slice of the network fires for any given request, so the running cost tracks the active number rather than the total. Open weights means Thinking Machines published the trained model itself, not just an API, so you can download it, run it on your own hardware, and modify it. Apache 2.0 is a permissive licence that allows commercial use, including inside closed products, as long as you keep the attribution notices.
Inkling Small actually beats its larger sibling on several tests, including Humanity’s Last Exam at 32 percent against 30, and GPQA Diamond at 89 against 87. It falls behind on agent-style tasks and on factual recall. The more practical difference is token efficiency: it averages around 24,000 output tokens per task, against 45,000 for DeepSeek V4 Flash and 78,000 for GPT-5.4 mini. Since you pay per token, that is a direct cost difference on top of the smaller size. It takes text, image and speech input, has a 256,000-token context window, and the weights are on Hugging Face. Artificial Analysis says no open model of the same size or smaller scores higher.
What is behind this. For two years the industry’s default answer to “how do we get a better model” was “make it bigger”. That is quietly ending. Inkling Small is one of several releases this month arguing that architecture, training data and token discipline buy you more than raw scale does. Thinking Machines has a specific reason to care: it positions its models as a base that customers fine-tune on their own data, and a 276-billion-parameter model is something a serious company can realistically fine-tune and host, where a 975-billion-parameter one is not. The strategy is to be the model you build on, not the model you rent.
What this means for you: if you use AI through a chat app, nothing changes, and you will probably never touch this model directly. If you are technical and have been priced out of self-hosting, the gap between “open model I can actually run” and “frontier model I have to rent” got noticeably narrower this week. The honest caveat is that 276 billion parameters is still far beyond a laptop, so open weights here means a company with GPUs, not a hobbyist with a MacBook. For everyone else, the useful signal is directional: smaller models keep catching up, and that keeps pushing prices down across the whole market.
Sources
OpenAI is now watermarking AI voices, and you can check a file yourself
Audio generated with GPT-Live through ChatGPT Voice and the API now carries an invisible SynthID watermark, and OpenAI's public verification tool can read it.
