Thinking Machines Releases Inkling, Its First Open-Weights Model Built to Be Fine-Tuned

Thinking Machines Lab released Inkling, its first in-house AI model, and put the full weights up for download. It is an open-weights, multimodal Mixture-of-Experts model that takes text, images, audio, and video as input, and the company is positioning it less as a finished product than as a base other teams retrain for their own work.

• Open weights, downloadable now. The full model weights are on Hugging Face in both standard and NVFP4 formats, the latter tuned for NVIDIA Blackwell hardware.

• Two sizes. Inkling runs 975 billion total parameters with 41 billion active per token. A preview variant, Inkling-Small, runs 276 billion total and 12 billion active.

• Built to be customized. Thinking Machines is steering users toward fine-tuning Inkling through Tinker, its model-customization platform, rather than shipping it as a one-size-fits-all system.

An open-weights MoE trained on text, images, audio, and video

Inkling is a Mixture-of-Experts transformer with 256 routed experts and two shared experts per layer, activating six routed experts per token. It was pretrained on 45 trillion tokens spanning text, images, audio, and video, then post-trained with large-scale reinforcement learning across more than 30 million rollouts, according to the model announcement. Training ran on NVIDIA GB300 NVL72 systems.

The model handles a context window of up to 1 million tokens and accepts multimodal input: images as 40x40 pixel patches, audio as spectrograms, and video. It also exposes a controllable "thinking effort" setting that lets users trade off how many tokens the model spends reasoning before it answers.

Inkling joins a run of open-weight releases aimed at teams that want to run and modify models on their own hardware. We covered Google's Gemma 4, which shipped under an Apache 2.0 license for local use without per-token fees.

Strong reasoning and coding scores, from a company that says it isn't the best

On benchmarks run at a high thinking-effort setting, Inkling reports 97.1% on AIME 2026, 87.2% on GPQA Diamond, and 77.6% on SWEBench Verified for agentic coding. With tool use, it scores 46.0% on Humanity's Last Exam. On multimodal tests it reports 73.5% on MMMU Pro for vision and 91.4% on VoiceBench for audio.

Thinking Machines is candid about where the model sits, stating that Inkling is "not the strongest overall model available today, open or closed." The pitch behind that framing, as TechCrunch reported, is a bet that "AI that organizations can adapt for themselves will outperform the one-size-fits-all models the biggest labs currently sell."

Access runs through Tinker, hosted APIs, and a free Playground

Teams can fine-tune Inkling on Tinker, which Thinking Machines is offering at a 50% limited-time discount with 64K and 256K context-length options. TechCrunch noted that Tinker, not the model itself, is where the company's revenue has to come from, through training, fine-tuning, and a cut of the hosting ecosystem around the weights.

For inference, the weights are hosted on TogetherAI, Fireworks, Modal, Databricks, and Baseten, and run on open tooling including SGLang, vLLM, and llama.cpp. A free Inkling Playground offers a chat interface with web search for a limited time.

Thinking Machines Lab, led by former OpenAI CTO Mira Murati, built Tinker as its first product before releasing any model of its own. Inkling is its first public model.

What Inkling gives creative-tech teams

For studios and product teams weighing where to run AI, an open-weights multimodal model with downloadable weights changes the calculation. Teams can fine-tune Inkling on their own data, run it on their own hardware without per-token API fees, and point one base at text, image, audio, and video tasks.

The constraint is scale. A 975-billion-parameter model needs serious compute to serve, and the smaller Inkling-Small preview points to where more accessible versions may land. For teams already working with downloadable model weights, Inkling adds a multimodal reasoning option with a permissive release and a customization platform built around it.