Inkling is the first open-weights AI model from Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati. The weights download free of charge under an Apache 2.0 license; only inference, on your own hardware or through partner APIs, carries a cost. Under the hood sits a mixture-of-experts design with 975 billion parameters, 41 billion of them active per query, plus a context window of one million tokens. It reads text, images and audio, then answers in text, code included.
Inkling takes in text, images and audio, and writes back text, from code to structured data. The architecture is a mixture of experts: of its 975 billion parameters, only 41 billion wake up for any given query, which keeps inference bills down. Training chewed through 45 trillion tokens of text, images, audio and video.
Its handiest trick day to day is the thinking-effort dial. You turn it down, answers land faster and cheaper; you turn it up, the model digs deeper. On one coding benchmark, Thinking Machines reports matching NVIDIA's Nemotron 3 Ultra with a third of the tokens. And rather than bluffing, Inkling was trained to flag what it is unsure about.
Thinking Machines states plainly that Inkling is not the strongest model on the market, open or closed. The wager sits elsewhere: ship a balanced foundation that each organization then fine-tunes on its own data through Tinker, the company's customization service, with 64K or 256K context options. At release it was the largest American open-weights model, a direct answer to DeepSeek V4, GLM 5.2 and Kimi K2.6, the names crowding the top of the LLM model rankings.
A few distinguishing marks noted at launch:
The quickest route is the Inkling Playground, a free browser chat put online at launch. For real work, several paths coexist (your desktop, for its part, will politely sit the full model out).
| Access route | What you get | Best for |
|---|---|---|
| Inkling Playground | browser-based chat, nothing to install | a first taste |
| Hugging Face | full BF16 weights, NVFP4 and GGUF variants | teams running a GPU cluster |
| Tinker | fine-tuning on your data, 64K or 256K context | domain customization |
| Partner APIs (Together AI, Fireworks, Baseten, Modal, Databricks) | hosted inference billed per use | plugging it into your apps |
Yes, Inkling's weights download free from Hugging Face under an Apache 2.0 license. The cost shifts to running it: either your own hardware or pay-per-use APIs at Together AI, Fireworks, Baseten, Modal or Databricks. Fine-tuning through Tinker is a paid service as well.
Strictly speaking, Inkling is open weights rather than fully open source: Apache 2.0 covers commercial use, modification and redistribution royalty-free, but the training data and training code stay private. That is the same standard set by Llama and DeepSeek releases.
Heavy iron: the BF16 checkpoint asks for 2 TB of pooled VRAM, the quantized NVFP4 version gets by on roughly 600 GB, and compressed GGUF builds circulate for llama.cpp. On modest machines, the realistic option is Inkling-Small with its 12 billion active parameters.
Both play in the big open-model league, DeepSeek V4 on the Chinese side, Inkling on the American one. Early comparisons hand some rivals the edge on terminal coding, yet Inkling remains the only one of the pack to hear audio and read images natively, with a one-million-token context.
Verdict: A leaderboard king, no; a base to build on, yes: if your team is after a multimodal model to fine-tune on its own data without leaning on a closed API, Inkling ranks among the most ambitious open foundations west of the Pacific.
