The open frontier is now plural
Nathan Lambert's Interconnects newsletter rounds up three open-weight releases that together define the current Pareto frontier of capability per parameter. The takeaway is that the open-weight ecosystem now contains multiple serious contenders, not a single leader, and the differences between them are as informative as their similarities.
Laguna S2.1, from Poolside, is a 118 billion total / 8 billion active parameter mixture-of-experts model built for agentic coding. It fits on a single DGX Spark, which is a meaningful operational detail for anyone who wants to run a frontier-tier model on hardware they own rather than rent. Poolside also adopted the OpenMDW license, which is an Apache 2.0-style free license with better legal backing for AI models specifically.
On long-horizon coding benchmarks, Laguna S2.1 holds its own against models several times its size. It scores 70.2% on Terminal-Bench 2.1 with thinking enabled, which puts it first among open, disclosed-size models on Poolside's compiled leaderboard, behind only larger or closed systems. On SWE-Bench Multilingual it scores 78.5%, topping the published table outright.
The clearest signal is DeepSWE v1.1, which still has real headroom. There, Laguna S2.1 scores 40.4% against DeepSeek-V4-Pro-Max's 9.0%, with roughly one-sixth the active parameters.
Inkling, a different shape
Inkling, from Thinking Machines, is a 975 billion total / 41 billion active MoE model that sits in a different size class. It is not yet as competitive on the published agentic benchmarks, scoring 63.8% on Terminal-Bench 2.1, but the architecture choices are different and the deployment story is different too. Inkling is positioned more as a research artifact than as a drop-in production model, which is the kind of bet a lab like Thinking Machines can make.
Kimi K3, the scale bet
Kimi K3, from Moonshot, is the biggest open model release in some time. At 2.8 trillion parameters with 50 billion active, it is a different systems class from the other two. The weights will be made available under a noncommercial license, requiring inference and fine-tuning providers to enter into a commercial agreement.
The trade-off is the right one to call out explicitly. Kimi K3 can score near the frontier on capability benchmarks while remaining hard to deploy, because serving a 2.8 trillion parameter sparse model is not a job most teams can take on. The weights are only the start; the surrounding infrastructure is what determines whether a model is actually usable.
What the Pareto frontier looks like in 2026
The interesting structural fact is that the open frontier now has three different operating points along the same Pareto curve. Laguna S2.1 buys efficiency with a small active parameter count. Inkling bets on research flexibility at medium scale. Kimi K3 bets on absolute capability at the cost of deployment complexity.
Closed frontier models like Claude Fable 5 and Kimi K3 itself (in its deployed form) still lead on several benchmarks, including Terminal-Bench 2.1 (88.0 and 88.3 respectively) and SWE-Bench Pro (80.3 for Claude Fable 5). The open story is not that the open models have caught up on the absolute leaderboard. It is that they have made the catch-up question moot by offering different points on the curve that closed models do not serve well.
For anyone choosing a model to build on, the answer is now contextual. If you need a self-hostable coding model that runs on a single workstation, Laguna S2.1. If you need raw capability and have the GPU budget, Kimi K3. If you want a research artifact with a thoughtful architecture, Inkling. The frontier is a curve, not a point.
What the open-closed gap looks like in practice
The closed frontier still wins on the absolute leaderboards. Claude Fable 5 and Kimi K3 in its deployed form lead on Terminal-Bench 2.1, SWE-Bench Pro, and the long-horizon agentic benchmarks that increasingly define what frontier means. The open-weight story is not that open models have caught the closed models. It is that the open models have made the catch-up question less interesting by offering different points on the curve that closed models do not serve well.
For a developer choosing a model, the practical decision tree now looks like this. If you need absolute capability and have the budget to pay for hosted inference, the closed frontier is still the right answer. If you need to self-host and your workloads can run on a single workstation, Laguna S2.1 is the right answer. If you need raw capability and have a serious GPU budget, Kimi K3 with its open weights is the right answer. If you want to study the architecture and are willing to deal with deployment complexity, Inkling is the right answer.
That decision tree did not exist a year ago. The fact that it exists now is the more important story than any single leaderboard position.