📊 Full opportunity report: What The First Inkling From Thinking Machines Means For AI Progress on ThorstenMeyerAI.com — validation score, market gap, and execution plan.
TL;DR
Thinking Machines has released its first foundation model, Inkling, openly available on Hugging Face under Apache 2.0. This marks a notable shift toward transparency and ownership in AI development, though some details remain uncertain.
Thinking Machines, a relatively new AI lab founded by former OpenAI CTOs, has released its first foundation model, Inkling, which is part of Murati’s Thinking Machines releases openly available on Hugging Face under the Apache 2.0 license. This is a significant development in the AI field, emphasizing transparency and ownership, and marks a departure from the industry norm of closed models.
Inkling is a 975-billion-parameter mixture-of-experts transformer supporting multimodal inputs (text, images, audio) with a 1-million-token context window. It was trained on 45 trillion tokens of diverse data, including text, images, audio, and video, and was pretrained using a hybrid optimizer on NVIDIA systems. For more on large language models, see our coverage of Murati’s Thinking Machines release of the Open-Weights 975B model. The model’s weights are now available openly, allowing organizations to download, modify, and deploy it independently.
Unlike typical releases, the model’s weights are under Apache 2.0, granting broad usage rights, but the training data and pipeline are not publicly disclosed. Additionally, reports suggest Thinking Machines maintains a separate Acceptable Use Policy (AUP) restricting surveillance, deception, and automated decision-making, which could conflict with the open-source licensing. This layered policy raises questions about the scope of open access and permissible modifications.
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-Source Model Release for AI Development
This release signals a shift toward greater transparency and control in AI development, allowing users to own and modify models without relying on API access. It challenges the industry trend of closed models, potentially accelerating innovation and democratization of AI technology. However, the layered restrictions via AUP may complicate open-source expectations, prompting industry debate about the balance between openness and responsible use.
HIWONDER Humanoid Robot with ChatGPT Multimodal AI Models AI Embodied Intelligent Vision Scene Voice Understanding 20DOF Educational Robot Kit Python Programming, TonyPi Advanced & RaspberryPi 5 4GB
- Powered by Raspberry Pi 5: High-performance AI vision robot
- Open-source Development Environment: Supports advanced AI robotics development
- ChatGPT Multimodal Integration: Enhanced human-machine interaction
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background on AI Model Releases and Industry Norms
Historically, most large AI models have been released with restricted access, often via APIs, to control usage and prevent misuse. Open-source releases have been rare and typically limited to smaller models or with significant restrictions. The recent trend has been toward proprietary models from major tech companies, with limited transparency about training data and methods.
Thinking Machines, founded 17 months ago by ex-OpenAI executives, has taken a different approach by releasing its weights openly, emphasizing ownership and transparency. This move comes amid ongoing debates about AI safety, data privacy, and the ethics of model deployment.
“We believe in empowering developers with ownership of the models they build upon, while maintaining responsible use through clear policies.”
— Thinking Machines spokesperson
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.
Uncertainties About Model Licensing and Usage Restrictions
It remains unclear how enforceable the separate Model Acceptable Use Policy (AUP) is, and whether it significantly limits the open-source potential of Inkling. The exact scope of restrictions and how they will be monitored or enforced by Thinking Machines is still unknown. Additionally, the impact of not releasing training data or pipeline details on transparency and reproducibility remains a concern among industry observers.
LLM Systems Engineering: Training and Building Large Language Models – Engineering AI Models Through Fine-Tuning, Continued Pretraining, and From-Scratch Development
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Next Steps for Industry Adoption and Independent Testing
Independent researchers and organizations are expected to evaluate Inkling’s performance across various benchmarks and real-world applications. Further scrutiny of the AUP and its enforcement will likely follow, alongside discussions about the implications of layered licensing on open-source models. The model’s adoption and adaptation by the broader AI community will reveal how this approach influences future releases.
AI Engineering: Building Applications with Foundation Models
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 notable for being openly available under the Apache 2.0 license, supporting multimodal inputs, and being trained on a diverse dataset with a large context window. Its open weights allow for independent modification and deployment, unlike most proprietary models.
Does the layered Acceptable Use Policy restrict the model’s openness?
Reports suggest there are restrictions via a separate AUP, which may limit certain uses such as surveillance or deceptive practices. The enforceability and scope of these restrictions are still unclear, raising questions about true openness.
What are the risks of releasing such a large model openly?
Risks include misuse for malicious purposes, potential safety concerns, and challenges in monitoring compliance with usage policies. The layered restrictions aim to mitigate some risks, but their effectiveness remains to be seen.
Will other companies follow Thinking Machines’ approach?
It is uncertain. While some may adopt more open models, many industry players remain cautious due to safety, legal, and competitive considerations. The impact of Inkling’s release will influence future openness strategies.
When will more details about the training data and pipeline be released?
There has been no official timeline. Transparency about training data and pipeline remains a concern, and further disclosures may depend on community feedback and regulatory developments.
Source: ThorstenMeyerAI.com
