On November 5, 2026, Mira Murati's Thinking Machines Lab unveiled its second-generation efficient model: a 594-billion-parameter mixture-of-experts with image input and strong coding, reasoning, and design scores, available first through Tinker with the open Apache 2.0 weights to follow after safety testing.
2026-11-07 · by Moe Ameen
Thinking Machines Lab, the AI lab founded by former OpenAI CTO Mira Murati, introduced Inkling 2 Small on November 5, 2026. It is a second-generation efficient model built as a mixture-of-experts with roughly 594 billion total parameters and about 17 billion active for any given token, so it runs far lighter than its headline size. The model takes text and image input through a vision encoder and responds in text, with up to a 1M-token context and a controllable "effort" dial that trades reasoning depth against speed. Notably, unlike the first Inkling and Inkling-Small, it does not accept audio input.
Rather than shipping open weights on day one as the earlier Inkling models did, Thinking Machines is making Inkling 2 Small available first through its Tinker platform with closed weights — you can run inference, fine-tune, and sample it in Tinker Playground now. The company said it plans to release the full weights under the permissive Apache 2.0 license after external safety testing and red-teaming, which it described as roughly a week out at announcement, with availability on additional platforms to follow. As of this writing the open-weights release has not yet shipped, so treat the date and availability as planned.
On the capability side, Thinking Machines reports strong benchmark results for the model's size — it scored around 31 on Artificial Analysis's Intelligence Index and posts competitive numbers on agentic coding, vision, knowledge-work, and design tasks. On one DesignArena website-building leaderboard snapshot it ranked as the top open-weights entry, a ranking the company flagged as pending the actual weights release. A distinctive trait is that it reads images and uses that vision to inspect and refine its own generated code, websites, games, and 3D scenes. The lab was candid about limits too: the model is less concise and less organized in chat than the best closed models, and early checkpoints showed alignment issues the team worked to reduce before launch.
The practical question for a creator today is what to actually do with Inkling 2 Small while its weights are still pending. The answer is use it through Tinker for what it's best at — reading a reference screenshot, drafting a script or caption pack, reasoning through a content plan — and let [Kompozy](/) handle everything the model can't. Inkling 2 Small writes words and reads images; it renders no media and posts nowhere.
Hand an Inkling 2 Small draft or an image-derived brief to Kompozy and it generates the finished formats the model stops short of: avatar-narrated [Persona Shorts](/glossary/persona-shorts) with burned-in captions, brand-exact [Carousels](/glossary/hyperframes) and quote cards, per-platform text, a blog article, and an email newsletter — every output held to one [Persona Brief](/glossary/persona-brief) so a week of posts sounds like a single brand. Then [Autopilot](/glossary/autopilot) schedules and publishes the whole set across eight social platforms plus blog and email, behind a per-post review gate. The model is the drafting and visual-reasoning front end; Kompozy is the engine that turns the plan into posts. For the full breakdown, see our [Inkling 2 Small deep-dive](/ai-tools/inkling-2-small), the [honest review](/reviews/inkling-2-small), and the [Inkling 2 Small alternative](/alternatives/inkling-2-small) page.
Inkling 2 Small is Thinking Machines Lab's second-generation efficient model, introduced on November 5, 2026. It is a mixture-of-experts — about 594B total parameters, roughly 17B active per token — that takes text and image input and responds in text, with up to a 1M-token context and a controllable reasoning-effort dial.
Not at launch. It is available through Thinking Machines' Tinker platform with closed weights for inference, fine-tuning, and sampling. The company plans to release the full weights under the Apache 2.0 license after external safety testing and red-teaming — described as roughly a week out at announcement — so treat the open release as planned, not yet shipped.
It is a newer, larger-but-still-efficient generation — about 594B total / 17B active, versus Inkling-Small's 276B/12B — with stronger agentic-coding, vision, and reasoning scores. The notable changes: it takes text and image input only (it dropped audio), and its weights are closed at launch rather than open on day one.
It can draft the text of a post and read a reference image, but it generates no images or video, burns no captions, and publishes nowhere. To turn its drafts into finished, scheduled posts you pair it with a generation and publishing engine like Kompozy, which renders the media and publishes across platforms.