📊 Full opportunity report: The AI Breakthroughs Hidden In Thinking Machines’ First Hints on ThorstenMeyerAI.com — validation score, market gap, and execution plan.
TL;DR
Thinking Machines has publicly released its first foundation model, Inkling, an open-weight, multimodal transformer with 975 billion parameters. The release emphasizes transparency and open licensing, but some restrictions may apply via a separate use policy. The model’s capabilities and limitations are now being scrutinized.
Thinking Machines has released its first foundation model, Inkling, a 975-billion-parameter, multimodal transformer, under the Apache 2.0 license. This marks a notable departure from typical proprietary releases, emphasizing open access and transparency in AI development. The full weights are now available on Hugging Face, enabling independent testing and deployment.
Inkling is a Mixture-of-Experts transformer supporting a 1-million-token context window, trained on 45 trillion tokens across text, images, audio, and video. It features a native multimodal input system, processing text, images, and audio jointly without external adapters. The model was trained with a hybrid optimizer on NVIDIA systems and underwent over 30 million reinforcement learning rollouts, with reasoning performance improving log-linearly during training.
While the weights are openly available, the training data and pipeline are not published, which is typical in industry but limits full reproducibility. Additionally, reports suggest that Thinking Machines maintains a separate Model Acceptable Use Policy (AUP) restricting surveillance, deception, and automated decision-making affecting individuals, raising questions about the true openness of the model’s use and licensing.
In benchmark testing, Inkling demonstrates strong performance in speech and safety benchmarks but shows middling results in pure language understanding tasks. The company explicitly states the model is not the strongest available today, prioritizing transparency and open access over raw performance.
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, Multimodal AI Release
The release of Inkling under an open license is a significant development in AI, enabling wider access, customization, and scrutiny of powerful foundation models. It challenges the norm of proprietary models by offering full weights and emphasizing transparency. However, the potential restrictions imposed by a separate use policy complicate the narrative, raising questions about the true openness and ethical use of the model. This move could influence industry standards and accelerate open AI research, but also underscores the importance of understanding licensing and usage restrictions.
Mastering OpenCode – Code with an Open-Source AI Agent in Your Terminal: Plan, Build, and Run Local Models with the OpenCode Agentic Coding Tool (Agentic Coding – Build Real Software with AI)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background on AI Model Releases and Industry Norms
Over recent years, most large AI models have been released as closed-source or with restricted access, limiting external validation and customization. The recent trend has been toward proprietary API-based models, with full weights often withheld. Open releases like Meta’s Llama 2 and EleutherAI’s models have aimed to democratize access, but often with licensing restrictions or incomplete transparency. Thinking Machines’ release of Inkling marks a notable shift, emphasizing full weight release under Apache 2.0, but with some caveats regarding usage policies. The model’s multimodal capabilities and scale position it among the most advanced open models to date.
“Thinking Machines’ release of Inkling is a milestone in transparency, offering open weights but with notable restrictions via a separate use policy.”
— Thorsten Meyer, Source
Multimodal AI Systems Engineering: Building Production Vision-Language Models, Document AI, and Cross-Modal Retrieval Pipelines (Production AI Engineering Series)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Remaining Questions About Inkling’s Use Restrictions
It is not yet confirmed how enforceable the separate Model Acceptable Use Policy is, or how it might impact commercial or research applications. The exact scope of restrictions and whether they could limit certain types of deployment remains unclear. Additionally, the transparency of the training data and pipeline has not been disclosed, leaving questions about reproducibility and bias unaddressed.
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.
Next Steps for Testing and Adoption of Inkling
Independent researchers and organizations will likely begin testing Inkling’s capabilities and limitations, focusing on benchmarks, safety, and ethical use. The full release of weights for the smaller Inkling-Small model is expected soon, providing additional avenues for evaluation. Monitoring how the separate use policy is enforced and how the community responds will be critical in assessing the model’s impact on open 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 open models?
Inkling is notable for its scale—975 billion parameters—and its native multimodal input capabilities, supporting text, images, and audio jointly. It is also released under Apache 2.0, making the weights freely available for modification and deployment.
Are there any restrictions on how I can use Inkling?
Yes, reports suggest that a separate Model Acceptable Use Policy may restrict certain applications, such as surveillance or automated decision-making affecting individuals. The enforceability and scope of these restrictions are not fully confirmed.
Why is the training data not published?
While the weights are open, the training data and pipeline remain proprietary, which is common in industry. This limits full reproducibility and transparency but does not prevent use of the model under the license.
Will Inkling be competitive with proprietary models?
In benchmarks, Inkling shows strong safety and speech capabilities but middling performance on some language understanding tasks. Its true competitiveness will depend on application-specific needs and how restrictions impact deployment.
Source: ThorstenMeyerAI.com
