📊 Full opportunity report: AI’s Quiet Revolution: Insights From Thinking Machines’ Inkling on ThorstenMeyerAI.com — validation score, market gap, and execution plan.
TL;DR
Thinking Machines has released Inkling, a 975-billion-parameter open-weight AI model, openly available on Hugging Face under Apache 2.0. The company emphasizes transparency about its performance and restrictions, marking a significant step in open AI development.
Thinking Machines has released its first foundation model, Inkling, a 975-billion-parameter transformer, openly available on Hugging Face under the Apache 2.0 license. This marks a notable shift towards transparency and ownership in AI development, contrasting with typical closed models and emphasizing the importance of owning and modifying AI models directly.
The Inkling model is a Mixture-of-Experts transformer supporting multimodal input—text, images, and audio—with a 1-million-token context window. It was pretrained on 45 trillion tokens of diverse data, including text, images, audio, and video. The model’s weights are fully released under Apache 2.0, allowing users to download, modify, and deploy independently.
Thinking Machines also introduced Inkling-Small, a 276-billion-parameter version that, thanks to an improved pre-training recipe, matches or exceeds the performance of its larger sibling on several benchmarks. The training process involved hybrid optimizers and over 30 million reinforcement learning rollouts, with some training data generated by open-weight models like Kimi K2.5.
The company explicitly stated that open weights are not the same as open source. The full training data and pipeline are not published, and there are reports that Thinking Machines maintains a separate Model Acceptable Use Policy (AUP) that restricts certain applications, such as surveillance and deception, which could conflict with the open licensing terms.
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 Model Release
The release of Inkling under an open license signifies a shift toward greater transparency and control for AI users, allowing organizations to own, modify, and deploy models without reliance on third-party APIs. This move addresses concerns about model access and control, especially after recent incidents where models were shut down by authorities. However, the potential restrictions imposed by the company’s AUP introduce questions about true openness and enforceability, which could influence how the model is adopted across sectors.
Yahboom K230 AI Development Board 1.6GHz High-performance chip/2.4-inch Display/Open Source Robot Maker Python, Supports AI Visual Recognition CanMV Sensor (with Heightened Bracket)
【Flagship performance, extremely fast response】Equipped with a 1.6GHz main frequency chip, the KPU computing power is 13.7 times…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background on Open AI Model Releases
Over recent years, most large AI models have been released with limited access, often through APIs, with the weights kept proprietary. Some organizations have begun to release open weights, but these typically come with restrictions or lack transparency about training data and pipelines. The recent launch of Inkling by Thinking Machines, a startup founded by former OpenAI CTO, marks a notable departure by releasing full weights openly and emphasizing ownership.
This approach contrasts with earlier models like GPT-3 or PaLM, which remain closed or partially open, often with licensing restrictions. The industry debate continues over the balance between openness, safety, and commercial interests, with Inkling’s release adding a new dimension to this ongoing discussion.
“We believe owning your model is essential for responsible AI deployment. Our transparency about performance and limitations is part of that commitment.”
— Thinking Machines spokesperson
Accelerate Everything with Tensor Cores: A Developer’s Guide to High-Performance AI, Efficient Training, and Scalable Models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Unresolved Questions About Inkling’s Use Policies
It remains unclear how enforceable the Model Acceptable Use Policy (AUP) is, and whether it significantly restricts the ways users can deploy or modify Inkling. The exact scope of restrictions and the potential for legal or practical conflicts with the Apache 2.0 license are still being evaluated. Additionally, the full training data and pipeline have not been disclosed, raising questions about transparency and reproducibility.
AI Engineering: Building Applications with Foundation Models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Next Steps for Adoption and Evaluation
Organizations and researchers will likely begin testing Inkling’s performance and compliance with the stated policies. Independent benchmarks and real-world deployments will clarify its capabilities and restrictions. Further disclosures from Thinking Machines about the AUP and training data are anticipated, along with potential updates or new models that expand on this open-weight approach.
Platform Engineering for Artificial Intelligence: Designing scalable infrastructure, data pipelines, and model lifecycle management for generative AI and agentic protocols (English Edition)
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 openly available under Apache 2.0, allowing users to own, modify, and deploy the model independently. It supports multimodal inputs and has a large context window, with performance claims supported by external benchmarks.
Are there restrictions on how I can use Inkling?
Yes, according to reports, Thinking Machines maintains an AUP that limits certain applications, such as surveillance or deception. The enforceability and scope of these restrictions are still being clarified.
Is the training data for Inkling publicly available?
No, the training data and full pipeline have not been published, which is typical for proprietary models, but it limits full transparency and reproducibility.
How might this release influence the AI industry?
It could set a precedent for more open yet controlled releases, balancing transparency with safety and legal considerations. It also raises questions about the true openness of models with layered policies.
Source: ThorstenMeyerAI.com
