Thinking Machines Lab debuts Inkling 2 Small, a MoE model with 594B total and 17B active parameters, and plans to release weights next week after safety testing (Thinking Machines Lab)

Inkling 2 Small: Why a 594B-Parameter AI Model That Only Uses 17B at a Time Changes the Economics of AI

By AI Worker Now Editorial · Published November 6, 2026 · Updated November 6, 2026

On November 6, 2026, Thinking Machines Lab introduced a new AI model called Inkling 2 Small. At first glance, the name sounds like a contradiction. The model has 594 billion total parameters, which makes it enormous. But it only uses 17 billion active parameters for each token it generates. That single design choice could reshape how AI is built, priced, and deployed.

The company also said it plans to release the model's weights next week, after safety testing. That means outside developers, companies, and researchers may soon be able to download the model and run it themselves. The combination of a huge total size, a small active size, and open weights is the real story here, not just another model launch.

What "Total" and "Active" Parameters Really Mean

Inkling 2 Small uses a mixture-of-experts (MoE) design. Think of a giant library with many specialized rooms. Each room holds a different kind of knowledge. When a question comes in, a router sends it to only a few rooms instead of opening the whole building. The total parameters measure the size of the entire library. The active parameters measure how many rooms are opened for each question.

For Inkling 2 Small, the library is vast: 594 billion parameters. But for each token, the model activates only 17 billion. That ratio is roughly 35 to 1. The model has a lot of stored capacity, but it does not pay the full computing cost on every step.

This matters because most AI costs come from active computation, not stored knowledge. A dense model with 594 billion parameters would need to run all of them for every token. That is slow and expensive. A sparse model with 17 billion active parameters can move much faster and cost much less per token, while still drawing on a much larger pool of learned patterns.

Why This Changes the Cost Curve

For businesses, the most important number is not total parameters. It is the cost to serve one request. Inference cost tracks closely with active parameters. A 17-billion-active model sits in a range that can run on a single high-end GPU, or a small cluster, depending on how the model is compressed and served. That is a very different hardware bill from a dense 594-billion-parameter model.

There is a catch. The full 594 billion parameters still have to be stored in memory. In standard 16-bit precision, that is hundreds of gigabytes of weights. Companies will need high-memory hardware or aggressive quantization to run it. But even with that memory cost, the compute cost per token is far lower than a dense model of the same total size. You pay to keep the library open, but you only pay to read a few books at a time.

This is the trade that makes MoE models so attractive. They let labs build models with enormous knowledge capacity without making every query expensive. For years, the biggest AI systems were locked behind expensive APIs because only large cloud providers could afford to serve them. Sparse designs like Inkling 2 Small start to break that link.

The Open-Weights Promise, With a Safety Gate

Thinking Machines Lab says it will release the weights next week after safety testing. That detail matters. Open weights mean outside teams can download the model, inspect it, fine-tune it on their own data, and run it inside their own systems. They are not limited to a vendor's API, a vendor's price, or a vendor's uptime.

The safety testing step is also notable. Releasing weights is not like releasing a closed API. Once weights are out, they can be copied and run anywhere. A short safety review before release is a sign that the lab is trying to balance openness with responsibility. It may also delay the release if testing finds problems. Companies planning to build on Inkling 2 Small should treat "next week" as a target, not a guarantee.

If the release happens as planned, it will put a very large sparse model into the hands of smaller teams. That is a shift in who gets to build with frontier-scale capacity. It does not mean the model is free to train, training a 594-billion-parameter model is still a massive expense. But it means the trained result can travel.

What Thinking Machines Lab Is Signaling

The name "Small" is interesting. It suggests this is not the largest model the lab intends to build. The active parameter count, 17 billion, is the number the lab wants you to notice. It is small enough to be practical, while the total parameter count is large enough to compete on knowledge and reasoning breadth.

Thinking Machines Lab is making a bet that efficiency and openness will matter more than raw size alone. That bet fits a broader trend in AI: the industry is moving away from "biggest model wins" and toward "best model per dollar." Sparse architectures, better routing, and open releases are all part of that shift.

There is also a competitive angle. If a lab can release a model that runs cheaply and can be fine-tuned by anyone, it builds an ecosystem. Developers who fine-tune Inkling 2 Small for a specific industry, law, medicine, finance, manufacturing, create value that the original lab does not have to build alone. Open weights turn users into contributors.

Practical Implications for Businesses

The first implication is lower cost for complex tasks. Many companies avoid large models because the per-token price adds up fast. A model with 17 billion active parameters can handle more traffic for the same budget. That makes advanced AI usable in high-volume workflows like customer support, document review, code assistance, and data extraction.

The second is private deployment. With open weights, a bank, hospital, or government agency can run the model inside its own network. Sensitive data does not have to leave the building. That is a major hurdle for regulated industries, and open weights clear it.

The third is fine-tuning on domain data. A company can take Inkling 2 Small and train it further on its own documents, tickets, or records. The result is a model that speaks the company's language and knows its products. Because the base model is sparse, that fine-tuned version is still affordable to serve.

The fourth is agents and multi-step workflows. AI agents often make many model calls in a loop: plan, act, check, revise. A slow or expensive model makes those loops impractical. A faster, cheaper model makes them viable. If Inkling 2 Small delivers on its active-parameter promise, it could be a strong engine for agentic software.

Hardware and Operations

Businesses should not assume this model is plug-and-play. Serving a 594-billion-parameter model, even a sparse one, requires careful engineering. Teams will need high-memory GPUs, efficient batching, and likely expert-parallel serving software. Cloud providers may offer it as a managed endpoint, but self-hosting will be a project, not a download-and-run experience.

The smart approach is to test both paths. Try a hosted version first to measure quality and latency. Then compare the cost of self-hosting against API pricing. For many companies, a hybrid model, hosted for spikes, self-hosted for steady private workloads, will make the most sense.

What It Means for Society

Cheaper, open AI models widen access. A small startup, a university lab, or a public-interest group can build on a model that would have been out of reach a few years ago. That can lead to more competition, more local innovation, and more tools built for languages and problems that big vendors ignore.

But openness also spreads risk. A model that can be downloaded can be misused. That is why the safety testing step matters. It will not solve every problem, but it sets a precedent: release weights, but do it carefully. Society will need more of these norms as large models become downloadable files rather than distant services.

Energy use is another factor. Sparse models reduce compute per query, which can lower energy use during inference. But the total memory footprint and the cost of training a 594-billion-parameter model are still large. Efficiency gains at inference do not erase the environmental cost of building the model in the first place.

Risks and Open Questions

There are real unknowns. A high total-to-active ratio does not guarantee quality. Routing can be uneven: some experts may be used far more than others, and the model may underperform on tasks that need broad, balanced knowledge. We do not yet have public benchmarks for Inkling 2 Small, so claims about its abilities should be treated with caution.

Serving complexity is another risk. MoE models are harder to optimize than dense models. Latency can vary depending on how requests are batched and which experts they hit. Teams should measure real-world performance, not just parameter counts.

Finally, the release date could slip. Safety testing is not a formality. If issues are found, the weights may come later or with restrictions. Companies should plan for that uncertainty and avoid building a product that depends on a single release date.

What to Watch Next

The actual weights and license. Open weights can mean many things. A permissive license, a research-only license, and a commercial license have very different business implications.
Real benchmarks. Look for independent tests on reasoning, coding, math, and long-context tasks, not just lab-reported numbers.
Community fine-tunes. The first wave of domain-specific versions will show how adaptable the model really is.
Hardware support. If major inference frameworks add fast support for Inkling 2 Small, adoption will speed up.
Larger siblings. The "Small" label hints at more models to come. Watch whether the lab scales the same sparse design up or down.

The bigger trend is clear. AI is moving toward models that are large in knowledge but small in active cost. Inkling 2 Small is a strong example of that direction: 594 billion parameters of capacity, 17 billion parameters of work per token, and a planned open release after safety testing. If it works as advertised, it will make advanced AI cheaper, more private, and more customizable, and it will push the rest of the industry to compete on efficiency, not just size.

TLDR: Thinking Machines Lab's Inkling 2 Small is a mixture-of-experts model with 594 billion total parameters but only 17 billion active per token. That design can deliver large-model knowledge at much lower inference cost. The lab plans to release the weights next week after safety testing. For businesses, this means cheaper complex AI, more private deployment options, and easier fine-tuning, though high memory needs and serving complexity remain real hurdles. The launch signals a broader shift toward efficient, open models over raw size alone.