Thinking Machines Inkling: Open Weights, Not Just Another Leaderboard Model
Thinking Machines Lab has released Inkling, its first open-weights model. The important part is not the 975B parameter headline. Thinking Machines is offering Inkling as a base for customization, with multimodal inputs, controllable reasoning effort, and a path to fine-tuning through Tinker.
That is a different proposition from launching another closed model and competing for the top general-purpose benchmark score. The lab says Inkling is not the strongest model available overall. Its bet is that a model developers can adapt is more useful than a slightly stronger model nobody can change.
What Inkling is
Inkling is a mixture-of-experts transformer with 975B total parameters and 41B active parameters per token. It accepts text, images, and audio, produces text, and has a 1M-token context window. OpenRouter lists it at $1 per million input tokens and $4.05 per million output tokens.
The model is available under the Apache 2.0 license through Hugging Face. Thinking Machines also positions Tinker as the practical route for post-training, fine-tuning, and reinforcement learning. That combination matters: the weights are available, but the intended workflow is not simply downloading a checkpoint and running inference locally. It is taking an open base and shaping it for a particular product or agent.
Inkling at a glance: LMRank score 8.5 · 975B total parameters · 41B active · 1M context · Apache 2.0
The tradeoff is customization, not leaderboard dominance
OpenRouter's current benchmark metadata puts Inkling's Intelligence Index at 40.7, Coding Index at 52.1, and Agentic Index at 32.3. Those numbers place it below the strongest general models in LMRank. The model should not be read as a cheaper replacement for every frontier API.
Its value is elsewhere:
- Controllable reasoning: applications can choose how much inference effort to spend instead of accepting one fixed thinking policy.
- Multimodal customization: image and audio inputs make it a more flexible starting point than a text-only open model.
- Long context: the 1M-token window leaves room for large codebases, document collections, and persistent agent state.
- Post-training access: developers can tune behavior, tool use, and domain knowledge rather than only adjusting a prompt.
The cost is operational complexity. A 975B-parameter checkpoint is not a casual single-GPU deployment, even with only 41B active parameters per token. Teams that want Inkling for production will need to think about serving infrastructure, quantization, batching, and the economics of fine-tuning. The Apache 2.0 license removes one restriction, but it does not remove the hardware bill.
Why Thinking Machines made this release
Thinking Machines Lab was founded by former OpenAI CTO Mira Murati around the idea that AI systems should be more adaptable to the people and organizations using them. Inkling is the first concrete model-level expression of that thesis.
The strategy also gives the lab a useful position in a crowded market. Closed providers compete on raw capability and managed reliability. Open-weight labs compete on access and cost. Thinking Machines is aiming at the layer between them: a capable foundation that a developer can specialize without building a pre-training stack from scratch.
That makes Inkling especially relevant for teams with a real reason to customize. A generic chatbot may get more value from a stronger hosted model. A company building a domain agent, a research assistant, or a multimodal workflow may value control over the last few points of benchmark performance.
Related: open-weight models · multimodal models · agentic models
The takeaway
Inkling is not Thinking Machines' attempt to claim the overall leaderboard. It is an open, multimodal, long-context base model aimed at developers who want to control the model's behavior and reasoning budget. At $1/$4.05 per million tokens through OpenRouter, the API price is competitive, but the bigger question is whether Tinker makes customization cheap and repeatable enough to justify the infrastructure.
For off-the-shelf intelligence, start with the higher-ranked models. For a model you can actually reshape, Inkling is one of the more interesting open-weight releases of the month.