[KongchangAI]
· 2 min read· 1,492 words

594B Open-Source MoE Model Inkling 2 Small: A Safe Path or a New Foundation?

594B Open-Source MoE Model Inkling 2 Small: A Safe Path or a New Foundation?

TML open-sources a 594B/17B multimodal MoE model under Apache 2.0, sparking debate over delayed open-sourcing.

Thinking Machines Lab has open-sourced Inkling 2 Small, a mixture-of-experts model with 594B total parameters, 17B activated parameters, and image input support, under the Apache 2.0 license, calling it the lab's strongest open-weight model. The article analyzes its so-called "safe path" of "closed first, open later": supporters say the closed-source window enables red-team testing and risk mitigation, while skeptics argue risks were already exposed via the API, the delay looks more like commercial pacing control, and the permissive Apache 2.0 license makes ex post constraints nearly impossible. Architecturally, the MoE design controls inference cost with 17B activated parameters, but the 594B total weights still require multi-GPU or high-memory deployment. The article concludes by discussing whether open source is truly narrowing the gap with the closed-source frontier and notes that the legitimacy of delayed open-sourcing depends on whether risk assessment methods are disclosed.

Thinking Machines Lab has open-sourced the weights of Inkling 2 Small under the Apache 2.0 license. It is a mixture-of-experts (MoE) model with 594B total parameters, 17B activated parameters, and support for image input. According to the official statement, this is the lab's most capable open-weight model to date.

Reddit discussion: Inkling 2 Small open source and the "safe path"

Just as interesting as the model itself is the release strategy behind it. TML calls this a "safe path to open weights": first release the model as closed source, then open the weights after some time. This approach raises a series of questions: what safety benefit does delayed open-sourcing actually buy? Once a 594B/17B model is released for free, will the gap between open and closed source continue to narrow, or will it instead give large labs a new foundation to build on?

The Logic and Controversy of the "Safe Path"

The core assumption behind TML's "closed first, open later" strategy is that during the closed-source phase, the model can undergo stricter red-team testing, abuse monitoring, and community feedback. Once risks have been fully identified and mitigated, the weights can then be opened. From a safety engineering perspective, this is not unreasonable—open weights mean anyone can download, fine-tune, and deploy the model, and once a model has dangerous capabilities, it is nearly impossible to收回 access afterward. The delay window gives developers a chance to "hit the brakes."

But the skepticism in the Reddit discussion is equally sharp: if the model was already exposed to a large number of users through an API during the closed-source period, then the safety risks already existed. Delayed open-sourcing looks more like commercial pacing control than a purely safety measure. Another view holds that真正有效的安全 comes from the model's capabilities and alignment quality themselves, not from the timing of a license release. In other words, if the model is safe enough, there is no essential difference between opening it on day one and on day one hundred; if it is not safe enough, delaying only pushes the problem further into the future.

The choice of Apache 2.0 makes the debate even more complicated. This is an extremely permissive license that allows commercial use, modification, and redistribution with almost no additional restrictions. For users, this means legal risk is greatly reduced; for safety advocates, it means that once the weights are out, any mitigation measures can only rely on community self-regulation.

The difference between open weights and open source: Strictly speaking, "open weights" only means that the model parameter files can be downloaded. Training data, training code, and data processing pipelines are often not disclosed, so it is not equivalent to open-source software in the usual sense. Once an open-weight model is released, anyone can fine-tune, merge, or remove alignment layers locally, and the model provider cannot remotely revoke access or impose usage monitoring as it can with an API service. This irreversibility is precisely the premise of the "safe path" discussion: the closed-source API phase preserves observability and the ability to intercept abusive behavior, but once the weights are out, safety constraints can only rely on the legal force of the license and users' self-discipline.

What 594B/17B MoE Means

Architecturally, Inkling 2 Small follows the currently mainstream MoE route: 594B total parameters provide capacity, while 17B activated parameters control inference cost. This design allows the model to maintain a large knowledge capacity while keeping the computation per forward pass close to that of a medium-sized dense model, significantly lowering deployment hardware requirements compared with a dense model of the same total parameter count.

The support for image input is also worth noting. The open-source community has had many options for text-only MoE models, but ultra-large MoE models with multimodal input capabilities and Apache 2.0 licensing remain scarce. For teams that need to build multimodal capabilities in-house, this reduces dependence on closed-source APIs.

However, 594B total parameters also bring practical barriers. Even though only 17B parameters are activated, the full weights still require considerable VRAM or memory to hold, and an ordinary single GPU can hardly accommodate them. Actual operation will most likely require multi-GPU parallelism or high-memory inference setups, which to some extent limits the imagination of "anyone can run it."

The basic mechanism of mixture-of-experts (MoE) architecture: An MoE model consists of multiple "expert" subnetworks and a router/gating network. During each forward pass, the router network activates only a small subset of the total experts (usually top-k, where k is much smaller than the total number of experts) based on the features of the input token, and the selected experts perform the computation and are weighted and aggregated. This is exactly where the separation between "total parameters" and "activated parameters" comes from: total parameters determine the upper limit of the model's knowledge capacity, while activated parameters determine the actual computation and memory access per inference. Compared with a dense model of the same total parameter count, MoE is more efficient in inference compute, but the cost is that all expert weights must reside in VRAM or memory. Therefore, it places higher demands on VRAM capacity (not just compute), which is the fundamental reason why a model with 594B total parameters is difficult to fit on a single GPU.

Is the Open-Source Frontier Closing the Gap or Feeding Closed Source?

This is the most tension-filled question in the discussion. Optimists believe that the open-weight frontier is rapidly approaching closed-source labs, and models like Inkling 2 Small allow small and medium-sized teams to access near-frontier capabilities, lowering the barrier to innovation.

Pessimists point to another cycle: large labs open-source their previous-generation models, gaining community ecosystem and reputation while shifting maintenance and deployment costs to users, while they continue to lead with stronger unpublished models. Open-source models become the "foundation" and "talent funnel" for closed-source models, and the gap does not necessarily narrow.

Both views have merit, but the key lies in the measurement standard. If we look at the absolute gap between "the strongest available open-source model" and "the strongest available closed-source model," open source is indeed catching up; if we look at whether the lab behind the open-source model still retains a significantly stronger internal version, the answer is often also yes. TML's own "closed first, open later" strategy is itself a microcosm of this structure.

Practical Deployment and License Considerations

For teams that want to get hands-on, the greatest value of Apache 2.0 lies in certainty: it can be used commercially, modified, and redistributed without worrying about additional terms. This makes it more popular than many "open source but restricted" licenses in enterprise internal deployment and secondary development scenarios.

On the hardware side, planning is needed for loading the weights of 594B total parameters, and multi-GPU inference or high-memory machines are more realistic starting points. Image input capability means that a multimodal processing pipeline needs to be reserved in the inference stack.

As for whether the "safe path" holds up, ultimately it depends on whether TML later discloses its risk assessment methods during the closed-source phase and the basis for its openness decisions. Without this transparency, the legitimacy of delayed open-sourcing is difficult to independently verify. For users, Apache 2.0 gives freedom, but it also hands safety responsibility over to them.

The substantive meaning of the Apache 2.0 license: Apache 2.0 is an OSI-approved open-source license that grants users a perpetual, worldwide, royalty-free copyright and patent license, allowing commercial use, modification, redistribution, and sublicensing, and does not require derivative works to be released under the same license (non-copyleft). Its main obligations are to retain copyright and license notices, mark modifications, and include a copy of the license when distributing. The patent grant clause it contains is an important difference from MIT/BSD, providing users with additional certainty at the patent level. It should be noted that the license constrains legal liability in distribution behavior and does not constitute a technical capability limitation or usage-scenario review.

One-line Summary

TML open-sources a 594B/17B multimodal MoE model under Apache 2.0, sparking debate over delayed open-sourcing.

Paragraph Summary

Thinking Machines Lab has open-sourced Inkling 2 Small, a mixture-of-experts model with 594B total parameters, 17B activated parameters, and image input support, under the Apache 2.0 license, calling it the lab's strongest open-weight model. The article focuses on its so-called "safe path" of "closed first, open later": supporters argue that the closed-source window allows for red-team testing and risk mitigation, while skeptics argue that risks were already exposed through the API, that the delay looks more like commercial pacing control, and that the permissive Apache 2.0 license makes ex post constraints nearly impossible. Architecturally, the MoE design controls inference cost with 17B activated parameters, but the 594B total weights still require multi-GPU or high-memory deployment. The article concludes by discussing whether open source is truly narrowing the gap with the closed-source frontier and notes that the legitimacy of delayed open-sourcing depends on whether risk assessment methods are disclosed.

#Inkling 2 Small#open-source MoE model#Apache 2.0#Thinking Machines Lab#open weights#multimodal model
Share:

Related articles