📊 Full opportunity report: Why Thinking Machines’ Inkling Matters For The Future Of AI on ThorstenMeyerAI.com — validation score, market gap, and execution plan.
TL;DR
Thinking Machines has publicly released Inkling, a large, open-weight multimodal AI model under Apache 2.0 license, marking a shift toward transparency. The model is not the strongest but exemplifies open access and honesty about capabilities.
Thinking Machines has publicly released its first foundation model, Inkling, under an open-source license, making its full weights available on Hugging Face. This move emphasizes transparency and ownership, contrasting with typical industry practices.
Inkling is a Mixture-of-Experts transformer with 975 billion total parameters and 41 billion active, supporting a 1-million-token context window. It was trained on 45 trillion tokens of text, images, audio, and video, and is multimodal, accepting text, images, and audio inputs without an encoder. The model’s weights are released under the Apache 2.0 license on Hugging Face, allowing download, modification, and commercial use.
In addition to the full-sized Inkling, a smaller version, Inkling-Small, with 276 billion total parameters and 12 billion active, was introduced, showing competitive performance on several benchmarks. The training process included hybrid optimization and reinforcement learning, with some training data generated by open-weight models like Kimi K2.5, highlighting transparency about the training pipeline.
However, the announcement also indicated that Thinking Machines maintains a separate Model Acceptable Use Policy (AUP), which restricts certain uses such as surveillance and deception, raising questions about the extent of openness and enforceability of these restrictions.
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 under an open license signifies a notable shift toward transparency and democratization in AI. Unlike many large models that remain closed or proprietary, Inkling’s open weights allow developers and researchers to inspect, modify, and deploy the model independently, fostering innovation and accountability.
However, the presence of a separate AUP introduces a layer of restrictions that could complicate open use, especially in sensitive domains. This move challenges the industry norm of open-sourcing models without additional restrictions, prompting discussions about the true nature of ‘openness’ in AI models and the balance between transparency and control.
multimodal AI development tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Industry Norms and the Significance of Transparency
Traditionally, leading AI models from companies like OpenAI and Google have been released with limited access, often as closed APIs or with restricted weights. Open-source models have generally been smaller or less capable, with full weights rarely shared openly. Recent efforts, such as Meta’s Llama and EleutherAI’s models, have pushed toward openness, but concerns about misuse and proprietary data remain.
Thinking Machines, founded by former OpenAI CTO, has built a reputation for transparency and rigorous testing. Its decision to release Inkling openly, alongside honest assessments of its capabilities, aligns with a broader movement toward open AI, but the layered restrictions via AUP highlight ongoing tensions between openness and control.
“We believe in open access to large models, but responsible use is also our priority. The AUP helps us ensure the model is used ethically.”
— Thinking Machines spokesperson
open-source AI models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Remaining Questions About Inkling’s Use and Restrictions
It is not yet clear how strictly the Model Acceptable Use Policy will be enforced or interpreted in practice. The extent to which restrictions will impact commercial or research applications remains to be seen. Additionally, the full training data and pipeline have not been disclosed, raising questions about reproducibility and transparency beyond the weights.
Further independent testing and real-world deployment will be necessary to assess the model’s capabilities and limitations fully.
Fine-Tuning Large Language Models: From Custom Datasets to High-Performance AI Models Using Modern Toolchains
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Next Steps for Adoption and Evaluation of Inkling
Expect ongoing benchmarking and testing by the AI community to validate Inkling’s performance across diverse tasks. Companies and researchers will evaluate the practicality of using the model under the AUP restrictions. Thinking Machines may release additional versions or details, and the industry will watch whether this approach influences other open-source initiatives.
Regulators and ethical bodies may also scrutinize the restrictions and licensing model, shaping future norms around open AI models.
large language model accessories
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 available under the Apache 2.0 license, with full weights released publicly, unlike most proprietary models. It is also multimodal, supporting text, images, and audio inputs without an encoder, and trained on a diverse dataset.
Are there restrictions on how Inkling can be used?
Yes, according to reports, Thinking Machines maintains a Model Acceptable Use Policy that restricts uses such as surveillance and deception. The enforceability and scope of these restrictions are still unclear.
Will the full training data for Inkling be released?
No, the training data and full pipeline have not been published, which limits full transparency and reproducibility of the model.
How does Inkling compare in performance to other models?
According to external benchmarks, Inkling performs strongly on safety and reasoning benchmarks but is mid-pack on some language understanding tasks. Its smaller variant, Inkling-Small, matches or exceeds some benchmarks of the larger model.
What are the implications of this open release for the AI industry?
This move could encourage more transparency and open collaboration but also raises questions about restrictions and responsible use. It may influence future open-source AI projects and licensing norms.
Source: ThorstenMeyerAI.com
