📊 Full opportunity report: The Future Of AI Might Be Hidden In Thinking Machines’ Inkling on ThorstenMeyerAI.com — validation score, market gap, and execution plan.
TL;DR
Thinking Machines has publicly released Inkling, a large multimodal AI model with open weights under Apache 2.0. The release highlights transparency but also raises questions about usage restrictions. The development signals a shift toward open models with clear licensing, but some details remain uncertain.
Thinking Machines has released the full weights of its new multimodal AI model, Inkling, under the open-source Apache 2.0 license. This move marks a rare instance of a major foundation model being openly available with transparent licensing, making it accessible for download, modification, and deployment by anyone. The release is significant because it challenges the industry norm of closed or restricted models, emphasizing openness and ownership.
Inkling is a Mixture-of-Experts transformer with 975 billion parameters and a 41-billion active parameter subset, supporting a one-million-token context window. It was trained on 45 trillion tokens across text, images, audio, and video, with a native multimodal input system that processes text, images, and audio jointly without an encoder. The model is designed for flexible deployment, with full weights available on Hugging Face under Apache 2.0, allowing users to fine-tune, inspect, and deploy independently.
While the weights are openly accessible, the company clarified that the training data and pipeline are not publicly released, and the weights are not open source in the strictest sense. Additionally, reports suggest that Thinking Machines maintains a separate Model Acceptable Use Policy (AUP) that restricts certain applications, such as surveillance and deceptive practices, which could limit the model’s open nature. This layered policy has raised questions about the scope of the open license and restrictions on use.
Benchmark results show Inkling performing strongly in safety and reasoning tasks, with notable scores in adversarial safety tests and speech recognition benchmarks. However, it is mid-tier in some reasoning benchmarks, such as Humanity’s Last Exam and SWE-bench Pro, and ranks around tenth in human web development evaluations among open models. The model’s safety features, licensing terms, and performance metrics are key points of interest for potential users and industry observers.
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 Foundation Model Release
The release of Inkling under an open license represents a notable shift toward transparency and ownership in the AI community. It offers developers and organizations the ability to customize and deploy the model without licensing fees or restrictions typical of proprietary models. This could accelerate innovation, foster more open research, and challenge the dominance of closed models from major corporations. However, the potential restrictions embedded in the separate AUP highlight ongoing debates about true openness and the balance between transparency and control in AI development.
For industries relying on AI for sensitive applications, such as security, healthcare, or public safety, the licensing and use restrictions become critical considerations. The model’s open weights provide a valuable resource, but users must carefully evaluate the legal and ethical implications of the layered policies. Overall, the development underscores a broader industry trend toward open models, but with nuanced limitations that will influence future adoption and regulation.
As an affiliate, we earn on qualifying purchases.
Industry Trends Toward Open AI Models and Recent Developments
Over the past year, several organizations have begun releasing large models with open weights, challenging the traditional proprietary approach. Notably, Meta released Llama 2 with open weights, and other startups have followed suit to promote transparency and democratization of AI technology. However, many of these models still come with restrictions or layered policies that limit their use in certain domains.
Thinking Machines, founded by former OpenAI CTO, has been known for its focus on safety and transparency. Its recent release of Inkling, a 975-billion-parameter multimodal model, marks a significant milestone. Unlike some open models that are fully open source, Inkling’s weights are licensed under Apache 2.0, but reports indicate the presence of a separate AUP restricting certain applications. This layered approach reflects a broader industry debate about balancing openness with responsible use.
Previous efforts have shown that open models can perform competitively on benchmarks, but concerns about misuse and ethical considerations remain. The Inkling release is a step toward addressing these issues by providing transparency about model capabilities and restrictions, although the full implications are still unfolding.
“We believe in responsible openness. Our Model Acceptable Use Policy ensures the model is used ethically while empowering the community with full access to weights.”
— Thinking Machines spokesperson
Unresolved Questions About Inkling’s Usage Restrictions
It remains unclear how strictly the separate AUP will be enforced and whether it will effectively limit certain applications despite the open weights. The exact scope, enforceability, and legal implications of the layered policy are still under discussion. Additionally, the full training data and pipeline have not been published, raising questions about reproducibility and transparency in the training process. The performance of Inkling on independent benchmarks beyond those reported by the company also remains to be verified.
Next Steps for Industry Adoption and Testing
Expect independent researchers and organizations to evaluate Inkling’s performance across various tasks and domains, including safety, reasoning, and multimodal capabilities. Further clarification on the AUP and licensing restrictions will likely emerge as users test the model’s practical applications. The company may also release additional documentation or updates on the training data and pipeline. Monitoring how the community adopts and adapts Inkling will be key to understanding its impact on the AI ecosystem.
Key Questions
Is Inkling truly open source?
No, the weights are licensed under Apache 2.0, but the training data and pipeline are not publicly available, and the company reportedly maintains a separate usage policy with restrictions.
What are the main capabilities of Inkling?
It supports multimodal input (text, images, audio), has a large context window of one million tokens, and performs well on safety and speech benchmarks. Its architecture is a 975-billion-parameter Mixture-of-Experts transformer.
Will the restrictions affect its adoption?
Potentially, yes. If the layered AUP limits certain applications, organizations will need to evaluate whether the model’s capabilities align with their ethical and legal standards.
When will more benchmarks or independent evaluations be available?
Likely in the coming months, as researchers and users begin testing Inkling across diverse tasks and domains.
Source: ThorstenMeyerAI.com
