📊 Full opportunity report: How Thinking Machines’ Inkling Could Foretell AI’s Next Evolution on ThorstenMeyerAI.com — validation score, market gap, and execution plan.
TL;DR
Thinking Machines has released Inkling, a large open-weight multimodal AI model with 975 billion parameters. The release emphasizes transparency and ownership, marking a notable development in AI’s evolution. Key questions remain about licensing policies and the model’s full capabilities.
Thinking Machines has released Inkling, a 975-billion-parameter multimodal AI model, openly available on Hugging Face under the Apache 2.0 license. This marks a significant shift in AI development, emphasizing transparency and ownership over licensing restrictions, and directly addresses the industry debate over model ownership versus rental.
Inkling is a Mixture-of-Experts transformer supporting a 1-million-token context window, trained on 45 trillion tokens across text, images, audio, and video. Its architecture routes tokens through 66 layers with 41 billion active parameters, and it was trained using a hybrid optimizer on NVIDIA GB300 systems. The full model weights are publicly available, enabling users to download, modify, and deploy independently.
Notably, the release includes a smaller variant, Inkling-Small, with 276 billion total parameters, which reportedly matches or exceeds the larger model on several benchmarks. The training process involved over 30 million reinforcement learning rollouts, and the model was trained on synthetic data generated by open-weight models, including Chinese models like Kimi K2.5.
While the weights are open under Apache 2.0, the company has reportedly implemented a separate Model Acceptable Use Policy (AUP) restricting certain applications such as surveillance, deception, and automated decision-making affecting individuals’ rights, raising questions about the true openness of the model’s use.
The weights came first: what Inkling actually signals
Mira Murati’s lab shipped its first foundation model — and the model isn’t the story. The order of operations is: full weights, Apache 2.0, day one, before any closed API. Plus a rare concession — the lab says it’s not the strongest model available, open or closed.
- AIME 2026 97.1%
- GPQA Diamond 87.2%
- MCP Atlas (Nemotron 44.7%) 74.1%
- VoiceBench · open-weight audio frontier 91.4%
- FORTRESS adversarial · best open 78.0%
- ForecastBench · calibration 61.1
- HLE text-only (GLM-5.2 40.1%) 29.7%
- SWE-bench Pro (GLM-5.2 62.1%) 54.3%
- Terminal-Bench 2.1 (GLM-5.2 82.7%) 63.8%
- SWE-bench Verified (Fable 5 95.0%) 77.6%
- Design Arena · 2nd open, behind GLM-5.2 ~10th
A 0.2 → 0.99 effort setting trades reasoning tokens against cost & latency, so you get a curve, not a point. On Terminal-Bench 2.1 it reportedly matches Nemotron 3 Ultra at ~⅓ the tokens. Peak score is a vanity metric when you serve millions of calls; the cost curve is what ships. (Bonus: its chain of thought compressed on its own during RL — nobody rewarded it; efficiency did.)
Pitched as the Western alternative to Chinese open weights (censorship-resistance training is the differentiator). But GLM-5.2 still wins on agentic/reasoning and Kimi K2.6 often on multimodal: best American open model, second in the open field. The irony — post-training was bootstrapped on synthetic data from Kimi K2.5.
BF16 needs ≥2 TB aggregate VRAM (8× B300 / 16× H200). NVFP4 still needs ≥600 GB. Not a workstation model — a 512 GB fleet falls just short. “Open” ≠ “runnable.” Mitigations: 1-bit GGUFs (~74% acc.), hosted eval routes, and Inkling-Small (12B active) — the release local-first builders actually want.
Open weights used to be a consolation prize. Inkling is a strategic open release — Apache 2.0, natively multimodal, honestly marketed, published complete on day one, optimized for deployment rather than headlines (the model isn’t the product; the fine-tuning platform is). It doesn’t need to win every benchmark for that to matter. The frontier is learning that owning the base beats renting the API — arriving now from the inside. For the sovereignty buyer: ① a real Western hedge against being switched off · ② verify the use policy before you build · ③ check the VRAM, then benchmark vs GLM-5.2 & Kimi K2.6 on your task.
Implications of Open-Weight Release for AI Ownership
This release underscores a shift toward greater transparency and ownership in AI development, allowing organizations to fully control and modify models without relying on API access. It challenges the industry norm of proprietary models and could accelerate innovation by enabling wider access and customization. However, the potential restrictions via the AUP complicate the narrative of open source, raising questions about enforceability and scope, especially for sensitive applications.
As an affiliate, we earn on qualifying purchases.
Industry Norms and the Shift Toward Open Models
Until now, most large language models have been offered through APIs or with closed weights, limiting direct access and modification. Recent debates have centered around the costs and benefits of owning versus renting AI models. The release of Inkling, with its open weights and transparent licensing, marks a notable departure from this trend. The model’s release follows a period of industry reflection on model ownership, safety, and ethical use, especially after incidents involving model shutdowns due to regulatory or political pressures.
Previous open models, such as Meta’s Llama 2, have faced similar debates over licensing and restrictions. Inkling’s approach, emphasizing transparency and open access, could influence future releases and industry standards.
“We believe in empowering developers and organizations with full ownership of their models, while also promoting responsible use through our AUP.”
— Thinking Machines spokesperson
Uncertainties Surrounding Licensing and Use Policies
It remains unclear how enforceable the separate Model Acceptable Use Policy (AUP) is, given that the weights are licensed under Apache 2.0, which imposes no restrictions. The scope of restrictions and how they are monitored or enforced in practice are still unknown. Additionally, the full capabilities of Inkling, especially in real-world applications, are still being evaluated, with benchmark results indicating strengths and weaknesses.
Next Steps for Adoption and Evaluation
Expect further independent testing of Inkling’s performance across diverse tasks and domains. Organizations will likely scrutinize the AUP’s enforceability and consider how to integrate the model responsibly. The company may also release more detailed documentation or updates on licensing and use restrictions, clarifying how the model can be deployed in sensitive applications. Industry observers will watch for how competitors respond to this open-weight approach.
Key Questions
What makes Inkling different from other large language models?
Inkling is a 975-billion-parameter multimodal model with open weights under Apache 2.0, supporting text, images, and audio, and is openly available for modification and deployment.
Does open weights mean the model is fully open source?
No, the weights are openly licensed under Apache 2.0, but the training data, pipeline, and potential use restrictions via an AUP are not publicly disclosed, which complicates the open source classification.
What are the potential risks of using Inkling?
Risks include uncertainty over enforceability of use restrictions, potential misuse if restrictions are bypassed, and unknowns about the model’s safety and biases in deployment.
How might Inkling influence the AI industry?
Its open-weight approach could set a new standard for transparency and ownership, encouraging other developers to release large models openly and fostering innovation.
Source: ThorstenMeyerAI.com
