📊 Full opportunity report: AI Innovation Alert: What Thinking Machines’ Inkling Suggests on ThorstenMeyerAI.com — validation score, market gap, and execution plan.
TL;DR
Thinking Machines has released Inkling, an open-weight, multimodal AI model with 975 billion parameters, openly available on Hugging Face under Apache 2.0. The release highlights transparency and honesty about its performance and licensing conditions.
Thinking Machines has officially released the full weights of its Inkling model on Hugging Face, making it openly accessible under the Apache 2.0 license. This move marks a notable departure from typical industry practices, emphasizing transparency and ownership for AI developers and researchers.
Inkling is a Mixture-of-Experts transformer with 975 billion parameters and a 41 billion active parameter subset. It supports a one-million-token context window and was trained on 45 trillion tokens, including text, images, audio, and video. The model is multimodal, accepting inputs in text, images, and audio, with a unique encoder-free design that processes these modalities jointly from scratch.
Unlike many recent models, Inkling’s weights are publicly available on Hugging Face under Apache 2.0, allowing users to download, modify, and deploy independently. The model’s training involved hybrid optimization techniques, and it underwent over 30 million reinforcement learning rollouts. Notably, some of the training data was generated by open-weight models, including Chinese models like Kimi K2.5.
However, the release also includes a Model Acceptable Use Policy (AUP) that restricts certain applications, such as surveillance and deception, raising questions about the scope of its openness. The model’s benchmarks show strong performance in speech and safety tasks but more modest results in text-only benchmarks. The full weights are available now, with further testing and evaluation ongoing.
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 for AI Development
This release signifies a shift towards greater transparency and ownership in AI development, allowing organizations to independently deploy and modify powerful models without relying on proprietary APIs. It also challenges industry norms by openly sharing weights while maintaining restrictions through a separate use policy, highlighting ongoing tensions between openness and control. For developers and policymakers, Inkling’s release underscores the importance of licensing clarity and ethical use policies in open AI models, especially given its capabilities across modalities and its potential for diverse applications.
AI Engineering: Building Applications with Foundation Models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Industry Norms and the Rise of Open-Weight Models
In recent years, most large AI models have been released as closed or API-only offerings, with weights kept proprietary to maintain control and monetize access. The recent trend towards open-sourcing models like Meta’s Llama or EleutherAI’s GPT variants has challenged this norm, emphasizing transparency and democratization. However, many open models are still accompanied by restrictions or lack full licensing clarity. Thinking Machines’ Inkling marks a notable development by openly releasing its weights under Apache 2.0, a license that permits modification and commercial use, but with an added layer of usage restrictions through a separate policy. This approach reflects ongoing debates about balancing openness with responsible use, especially as models grow larger and more capable.
“We believe in empowering developers with ownership and transparency, but also recognize the importance of responsible use through our policy restrictions.”
— Thinking Machines spokesperson
AI Engineering: Building Applications with Foundation Models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Unresolved Questions About Inkling’s Use Policy and Data
It remains unclear how enforceable the separate Model Acceptable Use Policy is, and whether it effectively limits misuse in practice. Additionally, the specifics of the training data and pipeline are not publicly disclosed, raising questions about data transparency and potential biases. The full performance of Inkling-Small and other variants is still under evaluation, and independent benchmarks are pending.
LM Studio for Beginners: Run Private AI Models on Your Own Computer — No Cloud, No Code, No Subscription
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Next Steps for Evaluation and Adoption
Researchers and organizations will likely conduct independent testing of Inkling’s performance across various benchmarks and applications. Further clarification on the use policy and licensing details is expected to be published. Industry observers will watch how the model is adopted, modified, and whether its restrictions influence responsible AI deployment. Additional updates on model improvements and training data transparency are anticipated in the coming months.
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 notable for being openly released under the Apache 2.0 license, with full weights available for download and modification. It supports multimodal inputs and has a large context window, distinguishing it from many proprietary models.
Does open weights mean the model is fully open source?
No. While the weights are open under Apache 2.0, the training data and pipeline are not publicly disclosed. Additionally, a separate use policy may impose restrictions on how the model can be used.
What are the potential risks of releasing such a large model openly?
Risks include misuse for malicious purposes, such as misinformation or surveillance, especially if restrictions are not effectively enforced. The balance between openness and responsible use remains a key concern.
How might this release impact the AI industry?
It could accelerate democratization and innovation by providing accessible, powerful models, but also prompts discussions on licensing, ethics, and responsible deployment.
When will more performance benchmarks and data transparency be available?
Further independent testing and detailed disclosures are expected in the coming months as the community evaluates Inkling’s capabilities and limitations.
Source: ThorstenMeyerAI.com
