📊 Full opportunity report: How Early Signs In Thinking Machines Could Alter AI’s Course on ThorstenMeyerAI.com — validation score, market gap, and execution plan.
TL;DR
Thinking Machines released Inkling, a 975-billion-parameter open-weight model, openly acknowledging it is not the strongest available. This signals potential shifts in AI development and ownership models, with ongoing questions about licensing and usage policies.
Thinking Machines has released its first foundation model, Inkling, openly available on Hugging Face under the Apache 2.0 license. This marks a significant moment in AI development, as the company explicitly states it is not the strongest model on the market, emphasizing transparency about its capabilities and limitations.
The Inkling model is a Mixture-of-Experts transformer with 975 billion parameters and a 66-layer decoder-only architecture. It supports a 1-million-token context window and was trained on 45 trillion tokens of diverse data, including text, images, audio, and video. Its multimodal input system processes audio as spectrograms and images as pixel patches, trained from scratch with a custom design.
The model’s weights are released openly under Apache 2.0 license, allowing users to download, modify, and deploy independently. However, the company maintains a separate Model Acceptable Use Policy (AUP), which restricts surveillance, deception, and automated decision-making affecting individuals’ rights. This layered policy raises questions about the true openness of the model and its intended use.
In benchmark testing, Inkling shows strengths in safety and speech-related tasks but lags behind in some text-only comprehension benchmarks. The company’s candid approach includes sharing external benchmark scores, though these are still to be independently verified.
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 Models for AI Ownership
The release of Inkling as an openly accessible model, coupled with an honest acknowledgment of its relative performance, signals a potential shift toward more transparent and owner-controlled AI systems. This could impact how AI models are developed, licensed, and used across industries, especially in sensitive domains like public safety and surveillance.
By openly sharing weights under a permissive license but layering restrictions through a separate policy, Thinking Machines exemplifies a nuanced approach that balances openness with control. This approach may influence future industry standards and raise important questions about the true nature of open-source AI, especially regarding enforceability and ethical considerations.
Generative AI for Software Development: Building Software Faster and More Effectively
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Recent Trends in AI Model Releases and Open Licensing
Over the past year, several AI labs have shifted toward releasing models with more transparency about capabilities and limitations. Notably, the recent launch of Inkling by Thinking Machines follows a pattern of openly sharing weights, contrasting with the industry trend of closed or partially open models. This approach responds to growing concerns about AI ownership, licensing, and the risks of unregulated deployment.
Historically, most foundation models have been released with proprietary licenses or limited access, often accompanied by closed training data and pipelines. The move toward open weights, as exemplified by Inkling, emphasizes user control and the potential for independent testing and modification, but also introduces new challenges related to misuse and compliance with layered policies.
„We believe in providing researchers and developers with open access to our models, but with responsible use policies in place.“
— Thinking Machines spokesperson
HIWONDER Humanoid Robot with ChatGPT Multimodal AI Models AI Embodied Intelligent Vision Scene Voice Understanding 20DOF Educational Robotic Kit Python Programming, TonyPi Pro & RaspberryPi 5 16GB
Al-Driven & Raspberry Pi Powered. TonyPi is a high-performance AI vision robot designed for AI education applications. It…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Uncertainties Surrounding Model Use and Licensing Enforcement
It remains unclear how effectively Thinking Machines’ separate AUP will be enforced, especially given the permissive Apache 2.0 license. The scope of restrictions and their legal enforceability in different jurisdictions are still to be tested. Additionally, the full training data and pipeline details have not been disclosed, raising questions about reproducibility and transparency.
Further, the actual impact of this open approach on AI safety, misuse, and industry standards is still unfolding, with ongoing debate about whether open weights truly enable responsible development or pose new risks.
How LLMs Actually Work: Large Language Models Explained: From Hardware to Hallucination, Tokens to Transformers
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Next Steps for Industry Adoption and Policy Development
Expect independent researchers and industry players to scrutinize Inkling’s performance and licensing policies further. There will likely be calls for transparency in training data and stricter enforcement of use policies. Additionally, other labs may adopt similar open-weight release strategies, influencing the broader AI ecosystem.
Regulators and policymakers might also examine the layered licensing approach to determine how best to regulate open-weight models without stifling innovation.
NVD RTX PRO 6000 Blackwell Professional Workstation Edition Graphics Card for AI, Design, Simulation, Engineering – 96GB DDR7 ECC Memory – 4th Gen RT/5th Gen Tensor Core GPU – OEM Packaging
[NVIDIA Blackwell Streaming Multiprocessor] The new SM features increased processing throughput, and new neural shaders that integrate neural…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What makes Inkling different from other foundation models?
Inkling is openly released under the Apache 2.0 license, allowing free download, modification, and deployment. It also explicitly states it is not the strongest model, emphasizing transparency about its capabilities and limitations.
Does open weights mean the model is fully open source?
No. While the weights are open under Apache 2.0, the training data, pipeline, and certain use restrictions are not publicly disclosed. Additionally, a separate use policy layers restrictions on top of the open license.
What are the potential risks of releasing open-weight models like Inkling?
Open weights can be misused for surveillance, deception, or automated decision-making that harms individuals. Enforcement of use policies and the transparency of training data are critical to mitigate these risks.
Will this approach influence other AI labs?
It is possible. The combination of open weights with responsible use policies may set a new standard, encouraging more transparent and owner-controlled AI development across the industry.
What should users consider before adopting Inkling?
Users should review the licensing terms, the separate use policy, and consider the model’s performance limitations in their specific applications, especially in sensitive domains.
Source: ThorstenMeyerAI.com
