The second generation of Inkling arrives with a detail that matters more for release strategy than for the raw spec sheet. Inkling 2 Small is a Mixture-of-Experts model with 594 billion total parameters and 17 billion active, able to accept image inputs and modulate compute effort. On performance, it far surpasses previous models on the Artificial Analysis Intelligence Index with a score of 31, with gains in agentic coding, vision, knowledge work, and reasoning. But the detail that counts for anyone looking at on-premises deployment is different: the model starts with closed weights on Tinker, and only after external safety testing and red-teaming will the full weights be published under an Apache 2.0 license.

This architecture has precise implications. In MoE models, not all 594 billion parameters run for every token: routing activates a subset, in this case 17 billion. This lowers computational cost per single inference compared to a dense model of comparable size, but it does not erase the memory cost: to serve the model locally you still need to host the entire set of experts, unless you use distribution or quantization techniques not specified in the note. The downloadable checkpoint, promised alongside the weight release, becomes a relevant object for anyone designing self-hosted infrastructure, because the footprint of stored parameters—not the number of active parameters—determines storage and, partly, memory bandwidth requirements.

The two-speed path chosen by Inkling signals a broader shift. The sequence 'closed API first, then opening after verification' is not an ideological concession to open source, but an attempt to reduce the risk of immediate uncontrolled use. For organizations that must respect data sovereignty constraints or operate in air-gapped environments, the moment the weights become available is when the model shifts from test object to an integrable component. Until then, the only access is mediated by Tinker, with the associated data transmission and latency constraints. The fact that checkpoints will be downloadable with the weight release reinforces the point: local fine-tuning is not an afterthought but a planned possibility.

There is a second structural signal. The jump from the previous generation—276 billion total and 12 billion active according to the comment in the source—shows that competition over open or semi-open models is no longer about the number of active parameters alone. It is about the combination of total size, routing efficiency, and release calendar. Anyone planning hardware capacity for inference must reason about both numbers: 17 billion active make per-token compute more manageable, but 594 billion total impose a set of accelerators and memory that few on-premises deployments can ignore. This is not a minor detail: models of this class, even with aggressive quantization, can require an amount of memory that depends on precision and distribution choices, a point the note does not resolve.

Variable effort support is another interesting lever: it allows compute intensity to be modulated according to the task, which in an on-premises context means containing inference costs without changing model. However, the source does not provide details on how this translates into latency or throughput.

Next week, when the weights are actually available, it will be possible to check whether the openness promises hold up against the real constraints of local infrastructures. For now Inkling 2 Small is above all a test of coherence: the race toward ever-larger models is not incompatible with a gradual opening path, but the entry cost for anyone who wants to bring it in-house remains tied to a number that the official announcement keeps in the background: 594B.