Thinking Machines Shrinks Inkling by Nearly Three Quarters and It Scores Higher

Fifteen days after Thinking Machines Lab put a 975-billion-parameter model on the internet for anyone to download, it published a second one roughly 72 percent smaller that beats the first on Humanity's Last Exam, SWE-bench Verified, GPQA Diamond and ARC-AGI-2. Model families are not supposed to work that way.

Inkling-Small landed on 30 July 2026, two weeks after Inkling, the first in-house model from the lab founded by former OpenAI chief technology officer Mira Murati. It is a Mixture-of-Experts transformer with 276 billion total parameters and 12 billion active per token, against Inkling's 975 billion total and 41 billion active. It keeps the 1-million-token context window, the native text, image and audio inputs and the adjustable reasoning effort of its larger sibling. It was trained on Nvidia GB300 NVL72 systems, and the weights ship under Apache 2.0 on Hugging Face.

The scores are the story. On the text-only Humanity's Last Exam, Inkling-Small posts 31.6 percent to Inkling's 29.7 percent. On SWE-bench Verified it hits 80.2 percent against 77.6. Terminal-Bench 2.1 reaches 64.7 percent versus 63.8. GPQA Diamond climbs to 89.5 from 87.2, ARC-AGI-2 to 40.1 from 36.5, IFBench to 82.2 from 79.8, and Toolathlon Verified to 54.4 from 45.5. Thinking Machines is unusually blunt about the asymmetry in its release post: 'With these improvements, Inkling-Small surpassed Inkling on reasoning and agentic coding benchmarks. Inkling maintains an advantage on knowledge coverage and factuality.'

That second sentence is doing heavy lifting. The regressions are steep. SimpleQA Verified collapses from Inkling's 43.9 percent to 20.6 percent. The Artificial Analysis Omniscience index falls from 2.1 to minus 9.0. Tau 3 Banking, an agentic customer-service benchmark, drops from 23.7 percent to 15.5. AIME 2026 slips from 97.1 to 95.5, and the audio evaluations tick down across the board: Audio MC 54.9 versus 56.6, VoiceBench 90.1 versus 91.4. Artificial Analysis put Inkling-Small at 40 on its Intelligence Index against Inkling's 41, and noted no open model of equal or smaller size scores higher.

Why the smaller model wins, and where it does not

The mechanism is not mysterious, and the lab spells it out. Inkling-Small started training after Inkling, which let engineers revise the pre-training data mix and the machine-learning recipe with the benefit of hindsight. An earlier checkpoint, Inkling-Small (preview), was then post-trained partly through on-policy distillation with Inkling as the teacher. From there the team kept scaling agentic coding reinforcement learning for another two weeks.

So this is not a story about sparsity magically buying you capability. It is a story about the difference between what a model knows and what a model does. Factual recall scales with total parameter count, because facts have to be stored somewhere, and 276 billion parameters simply hold fewer of them than 975 billion. Reasoning and tool use are procedural, and procedure transfers through distillation and reinforcement learning far more cheaply than memorised world knowledge does. The halving of SimpleQA is the price of admission; the gain on SWE-bench is what the extra RL bought. A student that watches a teacher work can learn the method without inheriting the library.

The efficiency case is the real product. The BF16 checkpoint needs 600 GB of aggregated VRAM, four Nvidia B300s or eight H200s. The NVFP4 quantised checkpoint drops that floor to 180 GB, which runs on a single B300 or two H200s. A 276-billion-parameter multimodal reasoning model that fits on one rented GPU instance is a different commercial object from a 2 TB model, and Thinking Machines knows it. The lab reports Inkling-Small averaging around 23,000 output tokens per task on GDPval-AA v2 against Inkling's 28,600, while scoring higher.

It is also worth being precise about what 'open weights' means here. Apache 2.0 is genuinely permissive, more so than the bespoke community licences Meta and others have used. But it governs the weights only. The 45-trillion-token training corpus behind the Inkling family is not published, the training code is not published, and Thinking Machines layers a separate Model Acceptable Use Policy over the licence that applies to the weights, related materials and any modified versions. This is open weights, not open source. It is also, notably, safety-devolved: the model card concludes Inkling-Small 'did not present risk of material uplift beyond what's already available in the open-weight ecosystem' and recommends deployers layer their own moderation, naming Llama Guard.

All of which sits on top of an unresolved business question. Thinking Machines has roughly 200 staff, no consumer product and a reported 50-billion-dollar funding round that was coming together last November and had stalled by January. Once weights are public, nobody who downloads them owes the lab anything. Revenue has to come from Tinker, the fine-tuning platform, and the company leans on that pitch explicitly: 'Tinker customers have seen firsthand that the right fine-tuned model can outperform closed models on a variety of tasks, and do so faster and cheaper.' The lab was equally candid at Inkling's launch that its flagship is 'not the strongest overall model available today, open or closed.' Well-rounded, cheap and adaptable is the entire wager.

Two things to watch. First, whether the factuality gap closes in the next generation, or whether Thinking Machines concedes recall to retrieval and search and stops optimising for it, which would be a meaningful architectural statement. Second, whether Tinker conversion actually materialises. Inkling-Small is a strong technical argument that distillation plus targeted RL beats brute parameter count for agentic work. It is not yet an argument that anyone will pay for it.