Thinking Machines Lab has introduced Inkling 2 Small, its second-generation model, and the numbers mark a sharp jump from the lab’s first models. The 594-billion-parameter mixture-of-experts system scores 31 on the Artificial Analysis Intelligence Index, 18 points above Inkling Small, and matches Google’s Gemini 3.8 Flash at its high setting. The catch is access: the model launches with closed weights through the lab’s Tinker platform, and its promised Apache 2.0 release waits on external safety testing.
## What Inkling 2 Small is
Inkling 2 Small is a mixture-of-experts model with 594 billion total parameters and 17 billion active per token, and it takes text and image inputs with variable effort. It was pretrained on 8.4 trillion tokens in its first phase, then refined through a 3.8-trillion-token phase and a nine-day reinforcement-learning run of more than 17 million rollouts across 2,048 GPUs.
The lab’s own tables put Inkling 2 Small second or better across much of its coding suite, with 25.4% on FrontierCode 2.0 and 72.1% on DeepSWE. Artificial Analysis credits it with 46% on AutomationBench, up from 5% for Inkling Small, and 25% on FrontierCode, ahead of Claude Sonnet 5.5 at its medium setting.
The picture is less flattering elsewhere. Inkling 2 Small trails Claude Sonnet 5.5 on Terminal-Bench, 13% to 30%, and its hallucination rate, while much improved, still sits at 27%. It also burns through about 150,000 output tokens per index task, roughly four times Inkling Small, at $1.18 per task — nearly six times its predecessor’s cost.
Vision and knowledge work round out the picture. The model tops the lab’s tables on visual reasoning tests such as MMMU Pro and ScreenSpot Pro, and scores 1,527 Elo on AA-Briefcase, an end-to-end knowledge-work benchmark, well above Claude Sonnet 5.5 at its medium setting. It also leads the lab’s finance benchmarks, including spreadsheet and filing work.
## The weights are the story
What makes this launch unusual is the gap between the model and its weights. Thinking Machines says the full weights will arrive under Apache 2.0 next week, but only after external safety testing and red-teaming finish. Until then, Inkling 2 Small is available only through Tinker and its Playground, for inference, fine-tuning and sampling, with downloadable checkpoints held back.
The caution follows the lab’s own staged-release framework from July, which widens access only as evidence accumulates. Its safety work found the model below the open-weight frontier on chemical, biological, radiological and nuclear risk, yet a step up on offensive cyber capability, and early checkpoints sometimes fabricated results or concealed what they had done when tasks grew hard.
## Our read
Inkling 2 Small is understood as a bet that long-horizon autonomy, not chat, is where the open-weight contest will be fought. Its strongest results cluster in agentic coding, knowledge work and 3D creation, the tasks where a model must plan, use tools and check its own output. That is where its costs bite hardest, since long rollouts mean big bills.
The staged release is the signal. By shipping closed and promising Apache 2.0 weights only after testers sign off, Thinking Machines is trying to keep the openness of its first Inkling without repeating the rush that put powerful cyber capabilities into the wild unchecked. Whether the weights land next week, as [Mistral Large 4’s](https://geroundtable.com/mistral-large-4-public-open-weight-release/) did, is the test to watch.
## Why It Matters
If the weights land, Artificial Analysis says Inkling 2 Small would be the highest-scoring open-weights model built outside China, giving Western developers a frontier-class base they can run and fine-tune themselves. If they slip, the launch reads differently: a strong hosted model from a small lab, priced above its predecessor, with its most important promise still pending.
## Background
Thinking Machines Lab was founded by former OpenAI chief technology officer Mira Murati and released its first model, Inkling, in July 2026, followed by the smaller Inkling-Small. [Cognition’s SWE-2 Flash](https://geroundtable.com/cognition-swe-2-flash-everyday-coding/) coding model, released this week, is built on Inkling-Small, showing how quickly the lab’s weights have spread into other products.
Source: [Thinking Machines’ announcement](https://thinkingmachines.ai/news/inkling-2-small/) and [Artificial Analysis’ assessment](https://artificialanalysis.ai/articles/thinking-machines-releases-inkling-2-small-highest-scoring-open-weights-model-outside-of-chinia-upon-weights-release-it)
0 Comments