📊 Full opportunity report: Can An Inkling From Thinking Machines Predict AI’s Path? on ThorstenMeyerAI.com — validation score, market gap, and execution plan.
TL;DR
Thinking Machines has released Inkling, an open-access AI model with 975 billion parameters, openly available on Hugging Face. While it offers transparency, it does not outperform the current top models, raising questions about its competitive edge and licensing restrictions.
Thinking Machines has publicly released the full weights of its new AI model, Inkling, on Hugging Face under the Apache 2.0 license. This move marks a notable shift towards open access in the AI community, especially given the model’s size and capabilities, and it directly addresses ongoing debates about proprietary versus open models.
Inkling 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 across text, images, audio, and video modalities. The model is multimodal, accepting text, images, and audio inputs, with training from scratch for its multimodal components.
The weights are openly available on Hugging Face under Apache 2.0, allowing users to download, modify, and deploy the model independently. However, the company has also reportedly implemented a separate Model Acceptable Use Policy that restricts surveillance, deception, and automated decision-making affecting individuals, raising questions about the true openness of the model’s use.
While Inkling demonstrates strong performance in certain benchmarks such as AIME 2026 (97.1%) and VoiceBench (91.4%), it does not lead in all areas, ranking mid to lower in some benchmarks. The model’s open weights and transparency are seen as a significant step, but its performance and licensing restrictions are points of contention.
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-Access Large Language Models
The release of Inkling under open weights represents a shift toward greater transparency and democratization in AI development, enabling wider research, customization, and deployment. However, the reported licensing restrictions and separate use policies highlight ongoing tensions between openness and control, impacting how organizations can leverage such models for sensitive applications.
This development matters because it influences the future landscape of AI innovation, potentially lowering barriers for smaller players and fostering competition, but also raising concerns around responsible use and enforceability of restrictions.
Accelerate Everything with Tensor Cores: A Developer’s Guide to High-Performance AI, Efficient Training, and Scalable Models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background on Open Models and Industry Norms
Until now, most large models have been proprietary, with access limited through APIs or licensing agreements. The release of open weights—such as those for Inkling—marks a departure from this norm, driven by a desire for transparency and community-driven innovation. Previous open models have often been smaller or less capable, making Inkling’s size and multimodal capabilities notable.
Thinking Machines, founded by former OpenAI CTO, has a history of building advanced models like ChatGPT, and its decision to release Inkling openly signals a strategic move to influence the AI ecosystem by providing accessible, modifiable models. The industry standard has often been to keep weights closed or partially open, citing concerns about misuse and safety, but Inkling’s release challenges this approach.
“We believe in providing our models openly to foster innovation, but with responsible use in mind.”
— Thinking Machines spokesperson
ESP32 Basic Starter Ai Chatbot Kit Development Board USB-C Dual Core Microcontroller Support AP/STA/AP+STA Compatible with Arduino IDE IoT for Beginners Engineers
【High Performance】The ESP32 module is equipped with a dual-core CPU and features a Type-C USB interface, with 44…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Unresolved Questions About Inkling’s Use Restrictions
It remains unclear how enforceable the reported Model Acceptable Use Policy is, and whether it effectively limits the model’s deployment in sensitive domains. The exact scope of restrictions and how they interact with the Apache 2.0 license are still being verified, raising questions about the true openness of the model.
Additionally, the performance of the upcoming Inkling-Small variant and its real-world applications are still under evaluation, and the full impact of the licensing restrictions on commercial use remains uncertain.
NVIDIA RTX PRO 6000 Blackwell Server Edition
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Next Steps for Adoption and Evaluation
Expect independent researchers and organizations to test Inkling’s capabilities across various domains, especially those concerned with licensing restrictions. The release will likely prompt further scrutiny of the use policy and possibly lead to clarifications or legal challenges.
In the coming months, more benchmarks and real-world deployment reports will clarify how Inkling compares to other models, and whether its open weights will accelerate innovation or face limitations due to restrictions.
Domain-Specific Small Language Models: Efficient AI for local deployment
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Is Inkling truly open source?
The weights are released under Apache 2.0, allowing download and modification, but reports suggest a separate use policy may restrict certain applications, complicating the notion of full open source.
What are the main capabilities of Inkling?
Inkling supports multimodal inputs (text, images, audio), has 975 billion parameters, a 1-million-token context window, and demonstrates strong performance in some benchmarks like AIME 2026 and VoiceBench.
How does Inkling compare to other models?
It ranks highly in safety and speech benchmarks but is mid-pack or behind in some language understanding tasks, indicating it is not the top-performing model overall.
What are the licensing restrictions on Inkling?
While the weights are openly available, reports suggest a separate Model Acceptable Use Policy restricts surveillance, deception, and automated decision-making, raising questions about the scope of open use.
What does this mean for AI development?
This release could democratize access to large models, fostering innovation, but also prompts debate over responsible use and the enforceability of restrictions layered on open weights.
Source: ThorstenMeyerAI.com
