Inkling 2 Small’s 71.8% Enterprise Score Tests the Limits of Benchmarks

The competitive significance of Inkling 2 Small is not simply that Thinking Machines has released another model. It is that a smaller system is claiming performance close to the frontier on work designed to resemble business operations. If that result holds up outside the benchmark, it could strengthen the case for deploying specialized, lower-cost models instead of relying on the largest available systems.

The harder question is whether one score can support that conclusion.

Thinking Machines Lab says in its launch announcement that Inkling 2 Small is a 594 billion parameter mixture-of-experts model, with 17 billion parameters active for each task. The company reports a 71.8% score on EnterpriseBench-CoreCraft, placing the model ahead of the other systems shown in its comparison table.

That architecture matters because the headline parameter count does not describe the amount of computation used for every response. A mixture-of-experts model can route a request through only part of its total network. In practical terms, the system is presented as having a large pool of capabilities while activating a smaller working set for each task.

For enterprise buyers, that distinction could become central. A model does not need to be the biggest system in the market if it can complete routine research, planning and operational tasks accurately enough at a lower serving cost. But “accurately enough” involves more than producing a plausible answer. Business agents must use tools correctly, follow instructions, satisfy all requirements and avoid errors that can propagate through workflows.

What CoreCraft is actually measuring

The benchmark behind the result is intended to test those capabilities in a controlled environment. Surge AI researchers describe CoreCraft in their paper as an enterprise simulation containing more than 2,500 entities and 23 tools.

That setup is more demanding than a conventional question and answer test. An agent operating in the simulated business must navigate a structured world, retrieve information, call tools and complete tasks against a rubric. The test is meant to approximate the coordination problems that arise when an AI system works across company data and software rather than responding to an isolated prompt.

The paper also reports a much lower baseline for frontier systems under strict evaluation. When every rubric criterion had to be met, the frontier models tested solved fewer than 35% of tasks. That result puts the reported 71.8% score in a striking light, at least within the conditions of the benchmark.

It also illustrates why the benchmark should not be treated as a direct measurement of general enterprise reliability. CoreCraft is a simulation, not a record of agents working inside a range of real companies with different processes, data quality, permissions and software environments. Its value depends on how faithfully those conditions represent the problems businesses actually face.

The research itself discusses limits on transfer and argues for realistic, diverse environments. That caveat is important. An agent can learn the patterns of one carefully designed world without possessing the broad operational judgment needed in unfamiliar ones. Strong performance may indicate genuine progress, but it may also reflect how closely the model's training and behavior fit the benchmark's structure.

The verification gap

The available public reception adds another qualification. BenchLM’s November 2026 leaderboard independently catalogs the 71.8% result, but says the only listed row is a provider self-report. It also warns that the evaluation harness details are not public and labels the result as display-only rather than an independently verified ranking.

That is the only critical response identified for this story, and it comes from BenchLM. No opposing company statement or independent reproduction was provided in the available sources. The warning therefore does not disprove Thinking Machines’ result. It defines how cautiously the result should be interpreted.

The difference between a reported score and a reproducible score is consequential. Without public harness details, outside researchers cannot readily determine whether the same prompts, tools, scoring rules and operating conditions would produce the same outcome. Buyers also cannot tell from the percentage alone how often the model made minor errors, how much human intervention was required or how performance changed when tasks were unfamiliar.

What the result could mean

Inkling 2 Small’s result still points toward a significant shift in model strategy. If compact or selectively activated systems can approach larger models on carefully defined business tasks, companies may have more options than choosing between general-purpose frontier models and conventional software automation. They could use smaller agents for bounded workflows, reserve larger systems for harder cases and evaluate each system against the cost of failure.

That future depends on evidence beyond a single benchmark. The next useful tests would include public evaluation procedures, independent reruns and results across different enterprise simulations. Real deployments would also need to report cost, latency, tool-use failures and the amount of supervision required.

For now, Inkling 2 Small has a credible and specific claim: 71.8% on EnterpriseBench-CoreCraft, with 17 billion active parameters in a 594 billion parameter mixture-of-experts design. The score suggests that smaller deployment footprints may no longer require an obvious sacrifice in benchmark capability.

It does not yet prove that the model is dependable across the untidy conditions of enterprise work. The future of business AI may belong to systems that are not merely powerful, but predictable outside the environment in which they were measured. That is the test Inkling 2 Small has not yet faced.