Thinking Machines Lab's second-generation efficient model — a 594B/17B mixture-of-experts with vision input, strong at code, design, and agentic reasoning.
Last verified · 2026-11-07 · by Moe Ameen
Inkling 2 Small is the second-generation efficient model from Thinking Machines Lab, the AI lab founded by former OpenAI CTO Mira Murati. It was introduced on November 5, 2026 and is a mixture-of-experts model with 594 billion total parameters and about 17 billion active for any given token, so it runs far lighter than its headline size. It takes text and image input through a vision encoder and responds in text, with up to a 1M-token context. Thinking Machines trained it natively in NVFP4 on NVIDIA GB300 NVL72 racks and attributes its gains to more efficient pretraining plus a single large, unified reinforcement-learning run that replaced the separate reasoning and agentic-RL stages used on the first Inkling.
At launch the model is available through the company's Tinker platform with closed weights — you can run inference, fine-tune, and sample it in Tinker Playground today. The full weights are planned for release under the permissive Apache 2.0 license after external safety testing and red-teaming, which the company described as roughly a week out at announcement; treat that date and the open-weights availability as planned, not yet shipped. On Artificial Analysis's Intelligence Index it scored around 31, and it posts strong numbers on agentic coding, vision, knowledge-work, and design benchmarks — on a DesignArena website leaderboard snapshot it ranked as the top open-weights entry pending its weights release.
What makes it interesting for builders is that it doesn't just write — it builds and checks visual work. It generates code, websites, small games, and 3D scenes, and uses its vision input to inspect and refine its own output. It also exposes a controllable "effort" dial to trade reasoning depth against speed and cost, following the pattern set by earlier Inkling models.
Two honest caveats. First, it is a text-output model: it reads images and reasons over them, but it generates no images, video, or audio of its own — and unlike the first Inkling and Inkling-Small, it does not accept audio input at all. Second, Thinking Machines notes it is less concise and less polished in chat than top closed models, and early checkpoints showed alignment issues the team worked to reduce. Like any model, its output needs checking before it ships.
Inkling 2 Small's distinctive trait is that it sees. Hand it a screenshot of a competitor's post, a rough mockup, or a reference frame and it reads the image and returns structured text — a hook set, a script, a caption angle, a content outline — then reasons and codes against it at the effort level you dial in. That makes it a sharp front end for turning visual inspiration into a written plan. What it will not do is render that plan into anything you can post: it outputs text only, generates no images or video, and publishes nowhere. And because its weights are still closed at launch (Tinker-only until the planned Apache 2.0 release), you're working through hosted inference, not a model on your own disk.
Kompozy is the layer that turns that written plan into finished, distributed content. Feed an Inkling 2 Small script or an image-derived brief into Kompozy and it generates the media the model can't: HeyGen [Persona Shorts](/glossary/persona-shorts) and avatar video with burned-in captions, face-locked Persona Photos and Persona Tweets, multi-slide [Carousels](/glossary/hyperframes) and Quote Graphics rendered pixel-exact through HyperFrames, plus full blog articles and newsletters — each held to your voice by the [Persona Brief](/glossary/persona-brief). Then [Autopilot](/glossary/autopilot) schedules and publishes the set across eight social platforms plus blog and email, behind a per-post review gate. The model reads the reference image and writes the plan; Kompozy builds the pixels and ships them.
Inkling 2 Small is Thinking Machines Lab's second-generation efficient model, introduced on November 5, 2026. It is a mixture-of-experts model — about 594B total parameters with 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 said it 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 that as planned, not yet shipped.
No. It reads images and reasons over them, but it outputs text only — it does not render images, video, or audio, and unlike the earlier Inkling models it does not accept audio input. To turn its text into finished posts you pair it with a generation and publishing engine like Kompozy.
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 benchmarks. The notable trade-off is that it takes text and image input only — it dropped the audio input the first Inkling line supported — and its weights are closed at launch pending the planned open release.
Inkling 2 Small writes the copy and can read a reference image, but it does not publish. Bring an Inkling 2 Small-drafted script or caption set into Kompozy to generate persona video, carousels, quote cards, blogs, or newsletters in your brand voice, then schedule and publish the set across TikTok, Reels, Shorts, X, LinkedIn, and more from one queue.