📊 Full opportunity report: The Hidden Messages In Thinking Machines’ Inkling About AI on ThorstenMeyerAI.com — validation score, market gap, and execution plan.
TL;DR
Thinking Machines publicly released Inkling’s 975-billion-parameter model weights under Apache 2.0, making it freely downloadable and modifiable. However, questions remain about licensing restrictions and transparency of training data, impacting its open-source status.
Thinking Machines has publicly released the full weights of its latest foundation model, Inkling, under the Apache 2.0 license, marking a significant move in the AI community. This release allows for free download, modification, and deployment, making Inkling accessible for a wide range of applications. However, the company’s accompanying policies and the model’s open-source claims are now subject to scrutiny, as questions about licensing restrictions and training data transparency emerge.
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, including text, images, audio, and video, with a custom training pipeline involving reinforcement learning and synthetic data generated by other open-weight models. The full model weights were released on Hugging Face under the Apache 2.0 license, enabling broad use and modification.
Despite the open licensing, reports suggest that Thinking Machines maintains a separate Model Acceptable Use Policy (AUP) that restricts certain types of use, such as surveillance and fully automated decision-making affecting individuals. This layered policy raises questions about the true openness of the model, as Apache 2.0 does not impose such restrictions by default. The company has not publicly verified or published the full AUP text, leaving some uncertainty about the scope of these restrictions.
In benchmark evaluations, Inkling demonstrated strong performance in safety and speech tasks, such as VoiceBench (91.4%) and FORTRESS (78.0%), but showed middling results in text-only reasoning benchmarks. The release is notable for its transparency about training methods and performance metrics, although independent validation is pending.
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 Weights with Usage Restrictions
The release of Inkling’s weights under Apache 2.0 provides the AI community with a powerful, freely accessible model, enabling innovation and customization outside of commercial API constraints. However, the potential restrictions imposed by the company’s separate AUP introduce ambiguity about the model’s true openness. For organizations and developers, understanding these restrictions is crucial before integrating Inkling into sensitive or regulated domains, such as public safety or surveillance.
This development underscores ongoing debates about what constitutes open source in AI—whether freely available weights suffice or if licensing and usage policies are equally important. It also highlights the industry’s evolving approach to balancing openness with responsible use, especially as models grow larger and more capable.
Samsung 43-Inch Class U8000H Series Crystal UHD, 4K, Smart TV, 2026 Model
- Crystal Processor: Enhances colors and sharpens details
- Free Streaming Content: Access 2,700+ free channels and content
- Motion Xcelerator: Delivers smooth motion in games and sports
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background on Model Releases and Open-Source AI
Over the past year, several AI labs have begun releasing large foundation models with varying degrees of openness. While some, like Meta and Stability AI, have released models under open licenses, concerns about misuse and proprietary training data have led to layered restrictions. Thinking Machines’ approach—releasing weights openly but potentially restricting use via a separate policy—reflects a broader industry trend toward balancing transparency with control. The company’s decision follows recent controversies over model misuse and shutdowns by governments, emphasizing the importance of licensing clarity.
Previously, open releases often included full datasets and training pipelines, but recent practices favor partial transparency to protect proprietary data and prevent misuse. The Inkling release, with its detailed benchmark results and transparent training methods, marks a notable step, although the ambiguity around licensing restrictions remains a point of contention.
“We are committed to responsible AI deployment and have policies in place to prevent misuse, which are layered on top of our open weights.”
— Thinking Machines spokesperson
Hiwonder AI Robotic Arm Kit Imitation Learning VLA Model Development Embodied AI Open Source 6-Axis Full Metal Programming Robot Arm with Magnetic Encoder Bus Servos & Tutorials, NexArm Standard Kit
- Embodied AI Robotic Arm: Industrial-grade metal, high-precision encoders
- Extended Reach and Payload: 500mm reach, 500g payload capacity
- Smooth and Precise Movements: Curve smoothing for jitter-free operation
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Unclear Scope of Usage Restrictions and Data Transparency
It remains unclear how broadly the Model Acceptable Use Policy (AUP) applies, whether it is enforceable, and if it significantly limits use cases. The full text of the AUP has not been publicly verified, raising questions about the extent of restrictions beyond the Apache 2.0 license. Additionally, the training data’s proprietary nature and the absence of a published training pipeline leave uncertainties about data transparency and reproducibility.
Ultimate Multimodal Transformer Models: Master LLMs, Vision Transformers, RAG, AI Agents, Fine-Tuning, and Multimodal AI Systems with PyTorch and Hugging Face (English Edition)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Next Steps for Validation and Policy Clarification
Independent researchers and organizations are expected to verify Inkling’s benchmark results and test the model’s capabilities in various domains. Clarification from Thinking Machines regarding the full scope of the AUP and its enforceability will be critical. Future releases may include more transparency on training data and policies, shaping how the community perceives and adopts this model.
Training Data for Machine Learning: Human Supervision from Annotation to Data Science
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Is Inkling truly open source?
While the weights are released under Apache 2.0, the presence of a separate Model Acceptable Use Policy suggests restrictions that complicate classifying it as fully open source.
What are the main restrictions on using Inkling?
Reports indicate restrictions against surveillance, deception, and automated decision-making affecting individuals, but the full scope of these restrictions is not publicly verified.
How does Inkling compare to other large models?
In benchmarks, Inkling performs strongly in safety and speech tasks but is middling in reasoning benchmarks, with performance metrics publicly reported for independent validation.
What does this mean for AI openness standards?
This release exemplifies the complex landscape of AI openness, where licensing and policies may diverge, impacting transparency and responsible use.
Source: ThorstenMeyerAI.com
