📊 Full opportunity report: Could Thinking Machines’ Inkling Be The Key To AI’s Next Leap? on ThorstenMeyerAI.com — validation score, market gap, and execution plan.
TL;DR
Thinking Machines has released Inkling, a large, open-weight, multimodal AI model with 975 billion parameters. The release emphasizes transparency but raises questions about licensing and use restrictions. Its impact on AI development remains to be seen.
Thinking Machines has officially released the full weights of its Inkling model under an open license, making it one of the largest openly available multimodal AI models to date. This move marks a significant development in AI, as it provides developers and researchers direct access to the model’s weights without restrictions on download, modification, or deployment, though questions about licensing and use restrictions remain.
The Inkling model is a Mixture-of-Experts transformer with 975 billion parameters, supporting a 1-million-token context window and trained on 45 trillion tokens across text, images, audio, and video. It features a multimodal input design, processing text, images, and audio natively, with a shared embedding space. The full weights were released on Hugging Face under Apache 2.0 license, allowing unrestricted use, modification, and commercial deployment, marking a departure from typical proprietary models.
However, the release was accompanied by a notable caveat: Thinking Machines reportedly maintains a separate Model Acceptable Use Policy (AUP) that restricts surveillance, deception, and automated decision-making affecting individuals. The company clarified that the Apache license does not impose such restrictions, but the layered policy raises questions about enforceability and scope, which potential users should verify before deploying the model.
In addition to Inkling, a smaller variant, Inkling-Small, with 276 billion parameters, demonstrated competitive performance on various benchmarks, thanks to an improved pre-training approach. The company also disclosed details about its training process, including hybrid optimization on NVIDIA systems and reinforcement learning with synthetic data generated by open-weight models like Kimi K2.5.
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
The release of Inkling’s full weights under an open license represents a major shift in AI accessibility, potentially enabling broader innovation and customization. This move could accelerate research, democratize access to large-scale models, and challenge proprietary AI ecosystems. However, the layered use restrictions and unclear enforceability of the AUP introduce risks and uncertainties for users aiming to build on this technology. The decision underscores ongoing debates about openness, control, and safety in AI development.
AI Systems Performance Engineering: Optimizing Model Training and Inference Workloads with GPUs, CUDA, and PyTorch
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background on Large-Scale Open AI Models and Licensing
Over the past few years, the AI community has seen a trend toward proprietary models with restricted access, often limiting research and commercial use. Recently, some organizations have begun releasing models with open weights, but typically with limited scope or accompanying restrictions. The release of Inkling’s weights under Apache 2.0 marks a notable exception, emphasizing transparency and user freedom. Historically, open-source models like GPT-2 and smaller variants have fueled innovation, but the scale and multimodal capabilities of Inkling set it apart as a potential game-changer.
Previously, most large models remained closed or partially open, citing safety, misuse, and intellectual property concerns. The explicit acknowledgment that Inkling is not the strongest model available today, yet is openly accessible, highlights a strategic focus on openness over competitive dominance, possibly setting a new precedent for the industry.
“We believe in open access to foundational AI models and want to empower developers and researchers to innovate freely.”
— Thinking Machines spokesperson
multimodal AI development kit
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Unclear Aspects of Licensing and Use Restrictions
While the weights are openly available under Apache 2.0, reports suggest that Thinking Machines maintains a separate Acceptable Use Policy (AUP) that could restrict certain applications, such as surveillance or automated decision-making. The exact scope, enforceability, and whether this layered policy will impact users’ rights remain unverified and are a matter of ongoing clarification.
Additionally, the full training data and pipeline have not been published, raising questions about reproducibility and transparency of the training process. The long-term implications of layered licensing and use restrictions in open models are still unfolding.
large language model GPU server
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Next Steps for Adoption and Industry Impact
Potential users and researchers will need to verify the details of the Acceptable Use Policy before deploying Inkling in sensitive applications. Further independent benchmarking and testing will clarify its real-world performance and safety profile. The company is expected to release additional details about the smaller Inkling-Small variant and its benchmarks soon. Industry observers will watch to see whether other organizations follow suit and how the layered licensing approach influences broader open-source AI development.
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
What makes Inkling different from other large language models?
Inkling is a 975-billion-parameter multimodal transformer with native support for text, images, and audio, and it is available under an open license, allowing unrestricted use and modification.
Does open weights mean the model is fully open source?
No. While the weights are released under Apache 2.0, the training data, pipeline, and potentially layered use restrictions are not fully disclosed, which limits full transparency and open-source status.
What are the risks of using Inkling given the layered restrictions?
The Acceptable Use Policy reportedly restricts certain applications like surveillance and automated decision-making. Users should verify these restrictions before deploying in sensitive domains, as enforceability is not yet clear.
Will Inkling’s open release impact AI safety and regulation?
The open release could accelerate innovation but also raises concerns about misuse and safety. Regulatory frameworks may need to adapt to balance openness with responsible use.
Source: ThorstenMeyerAI.com
