Thinking Machines Releases Inkling as a Customizable Open-Weights Model

Thinking Machines Lab has released Inkling, its first open-weights model. Confidence level: confirmed. The official launch page, model card, Hugging Face page, and third-party availability notes all point to the same core change: Inkling is meant for developers who want to customize a large multimodal base model, not just call another closed chatbot.

What changed

Thinking Machines says Inkling is a Mixture-of-Experts transformer with 975B total parameters and 41B active parameters. The company says it supports up to a 1M-token context window, accepts text, image, and audio input, and was pretrained on text, images, audio, and video.

The release matters because the weights are available through Hugging Face under Apache 2.0, while fine-tuning and API access are available through Tinker. Databricks also says Inkling is available through Unity AI Gateway for agent and application workflows.

Why this is early

This is early because the model launched today, July 15, 2026, and distribution pages are still fresh. The strongest evidence is official: Thinking Machines published the launch post and model card, and the Hugging Face repository shows the model page, license tag, file metadata, and last-modified timestamp.

The independent context is also useful but narrower. TechCrunch covered the launch, and Databricks published a day-zero availability note. LinkLoot treats those as corroboration for availability and market context, not as replacements for the official model card.

Key takeaways

• Inkling is an open-weights multimodal model from Thinking Machines Lab, founded by former OpenAI CTO Mira Murati.
• Thinking Machines says the flagship model has 975B total parameters, 41B active parameters, and up to a 1M-token context window.
• The model is positioned for customization, fine-tuning, agentic coding, tool use, chatbots, and retrieval-augmented applications.
• Hugging Face lists the model with an Apache 2.0 license, safetensors files, MoE tags, image-text and audio-text capabilities, and U.S. region metadata.
• The model card says teams should add application-layer safeguards and human review for high-stakes use.

Availability and access

Developers can inspect the model page on Hugging Face now. The model card says weights are available there and that API access is available through Tinker and third-party inference providers. The same card says local deployment requires serious infrastructure: the BF16 checkpoint needs at least 2 TB of aggregate GPU memory, while the quantized NVFP4 route lowers that to at least 600 GB.

That hardware note should drive the first decision. If you want to evaluate Inkling as a model family, start with Tinker, Databricks, or another hosted route. If you want to self-host, check the checkpoint format, inference framework support, GPU memory, license obligations, and safety layer before planning a deployment.

Practical LinkLoot angle

Inkling is not a "swap it into every prompt" release. The useful angle is customization. Teams that already fine-tune open models for internal workflows should compare Inkling against GLM, Kimi, DeepSeek, Nemotron, Qwen, and current closed models on their own agent tasks, not only public leaderboards.

Start with one bounded workflow: a coding agent that touches a known repository, a multimodal document task, or a retrieval workflow with measurable failure cases. Track cost, latency, instruction following, refusal behavior, hallucination rate, and how well fine-tuning changes the target behavior.

For broader agent tooling decisions, pair this evaluation with LinkLoot's AI agent tools guide.

What to verify before you act

• Confirm that the Hugging Face repository, model card, and license match your intended use.
• Check whether you can use Tinker, Databricks, or another provider in your region and account tier.
• Validate hardware needs before planning local deployment; the model card names 2 TB BF16 and 600 GB NVFP4 aggregate VRAM paths.
• Run your own safety tests for role-play, indirect prompts, multimodal inputs, and application-specific abuse cases.
• Compare Inkling on your workflow against both open-weights alternatives and closed frontier models before changing production defaults.

Source check

Confirmed by: Thinking Machines Lab's launch post names Inkling, describes the model family, and gives the 975B total parameter, 41B active parameter, 1M-context, multimodal, open-weights, and Tinker fine-tuning claims. The model card confirms intended uses, Apache 2.0 licensing, modalities, deployment paths, safety evaluation scope, and hardware requirements. Hugging Face confirms the public model page, tags, safetensors metadata, license tag, and repository status.

Independent context: Databricks confirms day-zero availability through Unity AI Gateway. TechCrunch covers the release as Thinking Machines Lab's first open model and frames it against the broader open-vs-closed model market. LinkLoot will treat new provider availability, benchmark audits, or model-card updates as update triggers.