📊 Full opportunity report: The Complex Meaning Behind Thinking Machines’ Inkling In AI on ThorstenMeyerAI.com — validation score, market gap, and execution plan.
TL;DR
Thinking Machines launched Inkling, a large open-weight multimodal AI model, openly available on Hugging Face. The release emphasizes transparency but raises questions about licensing and use restrictions.
Thinking Machines has officially released its first foundation model, Inkling, openly on Hugging Face under the Apache 2.0 license. This marks a notable departure from typical industry practice, emphasizing transparency and openness in AI model distribution, and directly addressing questions about the cost and control of owning AI models.
Inkling is a 975-billion-parameter multimodal transformer supporting text, images, and audio inputs, with a 1-million-token context window. It was trained on 45 trillion tokens across various data types, using a hybrid optimizer and over 30 million reinforcement learning rollouts. The model is accessible via Hugging Face, with full weights released under the Apache 2.0 license, allowing download, modification, and commercial use.
However, the release also includes a separate Model Acceptable Use Policy (AUP), which reportedly restricts certain uses such as surveillance, deception, and automated decision-making affecting individual rights. This policy raises questions about the true openness of the model, as Apache 2.0 licensing does not impose such restrictions. The model’s performance on various benchmarks shows strengths in safety and audio tasks but middling results in text-only tasks.
Thinking Machines claims the release prioritizes transparency and control, contrasting with typical industry practices of closed models or API-only access. The company emphasizes that users can inspect, fine-tune, and deploy Inkling independently, which is especially relevant following recent incidents where models were shut down due to external directives.
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 and Licensing
This release signifies a shift toward more transparent and controllable AI models, potentially empowering organizations to own and operate large models without reliance on third-party APIs. It also highlights ongoing tensions between open licensing and use restrictions, which could impact how the AI community perceives and adopts such models. For industries concerned with data privacy, security, and compliance, the combination of open weights with a restrictive AUP may influence future licensing and deployment strategies.
Nevertheless, the model’s open distribution raises questions about the balance between transparency and control, especially given the layered restrictions in the AUP. The industry will watch whether this approach becomes a model for future open AI releases or remains an isolated case.
multimodal AI model on Hugging Face
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background on Open-Weight AI Model Releases
In recent years, AI companies have often limited access to their large models through APIs or closed-source distributions, citing safety and commercial concerns. Open-source releases, when they occur, typically involve sharing weights under licenses like Apache 2.0 or MIT, but often with accompanying restrictions or lack of transparency about training data and safety measures.
Thinking Machines, founded by former OpenAI CTO, is notable for its focus on transparency and control. Its previous work includes models like GPT-3 and ChatGPT, but this is its first major open-weight release. The model’s release follows a broader industry trend toward more open models, driven by community demands for transparency and the desire to democratize AI development.
Prior to Inkling, most large models remained proprietary or API-restricted, with few exceptions such as Meta’s Llama 2 or EleutherAI’s GPT-NeoX. The release of Inkling under Apache 2.0, coupled with a potentially restrictive AUP, introduces a new approach that balances openness with control.
“Our goal was to provide a fully open-weight model that users can own and modify, while ensuring responsible use through our Acceptable Use Policy.”
— Thinking Machines spokesperson
open-weight AI models for commercial use
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Unresolved Questions About Licensing and Use Restrictions
It remains unclear how enforceable the Acceptable Use Policy (AUP) is, and whether it effectively limits the open-source nature of the weights. The exact scope of restrictions, especially in legal or commercial contexts, has not been independently verified. Additionally, the impact of these restrictions on the broader AI community’s perception of openness is still unfolding.
Further clarification is needed on how the AUP interacts with the Apache 2.0 license, and whether users can freely modify and deploy the model without risk of violating the policy.
AI model fine-tuning toolkit
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Next Steps for Adoption and Oversight
Industry observers will monitor how organizations adopt Inkling, especially regarding compliance with the AUP. Independent testing and benchmarking are expected to continue, with some groups likely to scrutinize the enforceability of restrictions. Future releases may clarify the legal boundaries and safety measures associated with open-weight models.
Additionally, discussions around licensing practices and community standards are anticipated to evolve, influencing how companies balance openness with responsible use.
large language model with audio and image support
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What makes Inkling different from other large language models?
Inkling is notable for being openly available under the Apache 2.0 license, allowing free download, modification, and commercial use, unlike many proprietary models. It is also multimodal, supporting text, images, and audio inputs.
Does the release mean the model is fully open source?
No. While the weights are openly available, the company reportedly maintains a separate Acceptable Use Policy that imposes restrictions on how the model can be used, which complicates the notion of full openness.
Why is the licensing and use policy important?
The licensing determines what users can legally do with the model, while restrictions in the AUP could limit or shape its deployment, especially in sensitive or regulated domains. Understanding these layers is crucial for responsible use.
What are the potential risks of open-weight models with restrictions?
There is a risk that restrictions could limit the model’s utility or lead to legal uncertainties. Additionally, layered restrictions might undermine trust in the openness of the release, affecting community adoption.
What happens next in the development and deployment of Inkling?
Expect ongoing benchmarking, scrutiny of licensing and restrictions, and possibly further clarifications from Thinking Machines. The industry will watch how organizations incorporate the model into their workflows and whether the layered approach becomes a new norm.
Source: ThorstenMeyerAI.com
