What the Numbers Show
Inkling 2 Small is a Mixture-of-Experts (MoE) model with 594 billion total parameters and 17 billion active per inference — a design that keeps compute costs manageable despite the model's size. It supports image inputs and variable effort, letting users trade response depth for speed depending on the task.
The headline claim comes from the Artificial Analysis Intelligence Index, where the company says Inkling 2 Small scored 31, which it describes as far surpassing the scores of its previous models. Thinking Machines attributes the gain to improvements "at every step" of training, including more efficient pre-training and a large, unified reinforcement learning run.
On coding benchmarks published in the announcement, Inkling 2 Small posted strong but not uniformly leading results against a field that includes GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017, Muse Spark, and Kimi K3. It scored 86.3% on SWEBench Pro v2 (Hard) — second-highest in the comparison, behind Kimi K3's 88.2% — and 72.1% on DeepSWE v1.1, where Muse Spark led at 77.0%. On FrontierCode v2.0 it reached 25.4%, second to Muse Spark's 28.5%, and it posted 81.6% on Vibe Code Bench v1.1.
Trained for Long-Horizon Agent Work
The company says the model's agentic abilities come from reinforcement learning on hundreds of thousands of long-horizon coding tasks, run across multiple agent harnesses and context-compaction strategies. Notably, Thinking Machines moved beyond simple unit-test rewards: many training tasks included an agentic RL grader that evaluated how the model approached a problem, not just whether the final code passed.
The team also tested early checkpoints in the wild, running them on OpenCode under the codename "Fledge-Alpha" to measure real-world performance. According to the announcement, the result is a model that follows instructions better than its predecessor, considers changes carefully, and verifies its own work without being asked.
Vision, Computer Use, and Interactive Demos
Inkling 2 Small extends the family beyond text. The company highlights demos in which the model used Three.js to build a full city simulator from scratch — generating every building, texture, and car procedurally with no pre-made assets — and wrote a working PC emulator in Python that boots a real Linux kernel, viewable in the browser. Another demo shows the model generating and revising scenes in Blender, including a hand-drawn valley with animated origami cranes.
Those demos underline the release's pitch: a model that can incorporate visual feedback and iteratively refine what it builds — websites, games, 3D designs — rather than producing one-shot output.
The Open-Weights Path
The staged release follows the approach Thinking Machines laid out in its earlier "safe path to open weights" essay. Inkling 2 Small is available today through Tinker, where users can try it in the playground and use it for fine-tuning and sampling. Checkpoints will become downloadable when the weights are released under Apache 2.0, planned for next week after external safety testing and red-teaming conclude.
The strategy echoes the company's first generation: the original Inkling flagship and the smaller Inkling-Small, a 276-billion-parameter open-weight model released earlier this year that drew significant attention for matching flagship performance at a fraction of the size. With Inkling 2 Small, the lab is doubling down on the same efficiency-first thesis — bigger total capacity, tiny active footprint.
Why It Matters
American labs have historically trailed Chinese competitors in shipping frontier-capable open weights, a gap that has narrowed as releases from Z.ai, DeepSeek, Tencent, and Moonshot land week after week. Thinking Machines is now the most prominent US entrant, and a second-generation release so soon after its first suggests the lab is iterating faster than many expected.
The caveats are the benchmark comparisons themselves. Thinking Machines chose the comparisons and labeled the "highest reported" scores in each row — and in several of those rows, Inkling 2 Small came second, not first. The Apache 2.0 release next week will be the real test: once developers can download and probe the weights independently, the marketing gives way to measurement. For now, the announcement positions Inkling 2 Small as evidence that the startup's model-product flywheel is turning — and that the open-weights frontier now has a genuinely transpacific contest.
Stay Ahead of AI
New open-weight models are landing every week from labs on three continents. Keep up with every release through our AI industry coverage.
Read more AI news →
