📊 Full opportunity report: The AI Signal We Missed: Insights From Thinking Machines’ Inkling on ThorstenMeyerAI.com — validation score, market gap, and execution plan.
TL;DR
Thinking Machines has released Inkling, a large open-weight multimodal model, openly stating it is not the top performer. The release emphasizes transparency about licensing and performance, raising questions about open-source claims.
Thinking Machines has released its first foundation model, Inkling, with full weights available on Hugging Face under Apache 2.0 license. The company openly states that Inkling is not the strongest model available, marking a shift toward transparency about model performance and licensing, which matters for developers and organizations considering open models.
Inkling is a 975-billion-parameter 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 across various media types, with the training process detailed as hybrid optimization on NVIDIA systems and over 30 million reinforcement learning rollouts. The model’s weights are openly available on Hugging Face, licensed under Apache 2.0, allowing free download, modification, and deployment. However, the company also reportedly enforces a separate Model Acceptable Use Policy restricting surveillance, deception, and automated decision-making, which could complicate open-source claims. Performance benchmarks show Inkling excels in speech and safety metrics but ranks mid-tier or lower on some text benchmarks, such as Humanity’s Last Exam and terminal reasoning tasks. The release includes a smaller variant, Inkling-Small, which matches or exceeds the larger model on several benchmarks thanks to improved training techniques. The transparency about not being the top model and the open weights are notable, but questions remain about the full training data and the scope of the use policy.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 Clarity
The release of Inkling’s full weights under an open license, combined with the admission that it is not the best model, signals a shift toward transparency and honest benchmarking in the AI community. It allows organizations to own, modify, and deploy the model independently, which could influence future open-source AI development. However, the reported existence of a separate use policy raises questions about the true openness of the model, especially for sensitive applications. This development could reshape expectations around open models, licensing, and responsible AI use, emphasizing the importance of clarity in licensing and restrictions for developers and regulators alike.
Large Language Models: The Hard Parts: Open Source AI Solutions for Common Pitfalls
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background on Open-Weight AI Model Releases
Over the past year, several AI labs have experimented with open-sourcing large models, but concerns about transparency, licensing, and safety restrictions have persisted. Previously, companies often released models with limited access or closed weights, citing safety and commercial reasons. Thinking Machines’ approach with Inkling marks a notable departure by openly sharing weights and being upfront about its performance relative to competitors. The industry norm has been to withhold full training data and impose restrictions through separate policies, which this release explicitly acknowledges. This context highlights a broader debate over what truly constitutes open-source AI and how transparency can be balanced with safety and commercial interests.“Our goal is to provide a genuinely open model that developers can own and adapt, with full transparency about its capabilities and limitations.”
— Thinking Machines spokesperson
Unanswered Questions About Licensing and Data Use
It remains unclear how the separate Model Acceptable Use Policy interacts with the open-source license, and whether restrictions could limit certain applications. The full training data and pipeline have not been disclosed, raising questions about data provenance and bias. Additionally, the actual enforceability of the use restrictions in practical deployment is yet to be tested and verified.Next Steps for Developers and Regulators
Developers will likely examine the model’s performance across diverse tasks and test the limits of the licensing restrictions. Independent researchers may scrutinize the training data and use policy for compliance and safety. Regulators and industry observers will monitor how the model’s open weights and restrictions influence standards for transparency, safety, and responsible AI deployment. Further disclosures from Thinking Machines regarding the training data and the enforceability of its use policy are anticipated.Key Questions
What makes Inkling different from other large language models?
Inkling is a 975-billion-parameter open-weight multimodal transformer, openly available under Apache 2.0, with a focus on transparency about its capabilities and limitations. Unlike many models, it openly shares its weights and admits it is not the top performer.
Does open weights mean the model is fully open source?
No. While the weights are under Apache 2.0, the training data and full training pipeline are not published. Additionally, a separate Model Acceptable Use Policy reportedly restricts certain applications, complicating the open-source claim.
Why does the performance of Inkling matter?
Its performance benchmarks show strengths in speech and safety, but it ranks mid-tier in some reasoning tasks. Its open availability and honest benchmarking could influence how future models are released and adopted, especially regarding transparency and licensing.
What are the risks of using a model with a restrictive use policy?
If the use policy imposes restrictions beyond the open license, it could limit certain applications, especially in sensitive domains like surveillance or automated decision-making. Users should review the policy carefully before deployment.
Source: ThorstenMeyerAI.com
