📊 Full opportunity report: The Critical Clues In Thinking Machines’ Inkling For AI Development on ThorstenMeyerAI.com — validation score, market gap, and execution plan.
TL;DR
Thinking Machines launched Inkling, a large, open-weight multimodal AI model, openly available on Hugging Face. The release emphasizes transparency but raises questions about licensing and use restrictions. The development signals a shift toward open models with clear limitations.
Thinking Machines has publicly released its latest foundation model, Inkling, under an Apache 2.0 license, making it freely downloadable and modifiable. This marks a significant moment in AI development, as the model’s weights are openly accessible, but the company also maintains a separate use policy restricting certain applications. The release underscores a shift toward transparency in large AI models, with important implications for developers and regulators alike.
Inkling is a 975-billion-parameter multimodal transformer supporting text, images, and audio inputs, with a 1-million-token context window. It was trained on 45 trillion tokens across diverse media, using a hybrid optimizer and over 30 million reinforcement learning rollouts, including synthetic data from open models such as Kimi K2.5. The model is not the strongest available, but its open weights under Apache 2.0 are notable for enabling independent inspection and modification.
Released on Hugging Face, the weights are accompanied by a model card stating the license, but reports suggest the company enforces a separate Model Acceptable Use Policy that restricts surveillance, deception, and automated decision-making affecting individuals. This layered policy raises questions about the true openness of the release, especially for sensitive applications. The model’s performance on benchmarks varies, with strengths in speech and safety but middling results in some language understanding tasks.
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.
Impact of Open-Weight Release with Use Restrictions
The release of Inkling as an openly accessible model under Apache 2.0 signals a new approach in AI development, emphasizing transparency and community engagement. However, the potential restrictions via a separate use policy highlight ongoing tensions between openness and control. For developers, this means the ability to fine-tune and deploy the model independently, but also the need to verify compliance with the company’s restrictions, especially in sensitive domains like surveillance or public safety. The move could influence industry standards on model licensing and responsible AI deployment, but also raises questions about enforcement and transparency.
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 Inkling’s Development and Release Strategy
Thinking Machines, founded 17 months ago by former OpenAI CTO, has built a reputation for transparency and technical rigor. Its prior work includes models with publicly available weights, but Inkling’s release under Apache 2.0 is notable for being one of the largest open models to date. The company’s approach contrasts with many industry players that restrict access via closed APIs. The model’s training involved diverse media and synthetic data, with a focus on safety benchmarks and performance metrics. The release follows recent debates over model openness, licensing, and responsible AI use, especially after incidents involving model shutdowns due to regulatory or political actions.
“We believe transparency and responsible use are compatible; our separate policy ensures the model is used ethically.”
— Thinking Machines spokesperson
multimodal AI development kits
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Unclear Aspects of Inkling’s Licensing and Use Restrictions
It remains unclear how strictly the separate Model Acceptable Use Policy will be enforced and whether it will impact independent developers’ ability to modify or deploy Inkling freely. The exact scope of restrictions, especially regarding surveillance, deception, and decision-making, has not been publicly verified. Additionally, the extent to which the policy aligns with the Apache 2.0 license remains ambiguous, raising questions about potential legal and ethical boundaries.
large language model GPUs
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Next Steps for Developers and Regulators
Developers will likely begin testing and fine-tuning Inkling, assessing its performance across various applications. Independent audits and benchmarks are expected to verify claims about safety and capabilities. Regulators and industry groups may scrutinize the licensing and use policies, especially in sensitive sectors. The company may also release further documentation clarifying restrictions and enforcement mechanisms. Monitoring how the community adopts and adapts the model will be crucial for understanding its impact on open AI development.
Secure AI Model Deployment: A Comprehensive Guide to Safely Delivering Machine Learning Systems in Production Environments
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 open weights under Apache 2.0, supporting text, images, and audio inputs. Its emphasis on transparency and its layered use policies distinguish it from many proprietary models.
Can I freely modify and commercialize Inkling?
While the weights are openly available under Apache 2.0, the company reportedly enforces a separate use policy that may restrict certain applications. Users should verify the policy and compliance requirements before deploying the model commercially.
Why is the layered use policy significant?
The policy could impose restrictions on surveillance, deception, or automated decision-making, potentially limiting the model’s use in sensitive areas despite its open weights. Its enforceability and scope are still being evaluated.
What are the performance strengths of Inkling?
Inkling shows strong results in safety benchmarks like FORTRESS (78%) and speech tasks such as VoiceBench (91.4%). Its language understanding benchmarks are middling, but its multimodal capabilities are notable.
Source: ThorstenMeyerAI.com
