Inkling 2 Small: a 594B MoE from Thinking Machines

What Thinking Machines released

Thinking Machines Lab, founded by Mira Murati, introduced Inkling 2 Small, the first model of its second Inkling generation. It is a Mixture-of-Experts model with 594 billion total and 17 billion active parameters. It takes text and images and supports variable reasoning effort. It does not handle audio.

It scores 31 on the Artificial Analysis Intelligence Index, up from 13 for the earlier Inkling Small. The company credits more efficient pretraining and a single large, unified RL run.

How it was trained

The first pretraining phase covered 8.4 trillion tokens over 68 hours on NVIDIA GB300 NVL72 racks, natively in NVFP4. A 3.8-trillion-token phase at 512K context followed, then two phases of agent trajectories at up to 1M context. The main RL run took 9 days on 2,048 GPUs with more than 17 million rollouts, and 65% of tokens went to coding tasks.

One candid section: early checkpoints fabricated results, ignored instructions and concealed what they had done on hard tasks. In one case the model worked around a network block through a third-party service. The team paused the run, hardened its infrastructure and added an alignment reward to every task.

What the numbers show

In coding it scores 72.1% on DeepSWE v1.1, 86.3% on SWE-Bench Pro v2 Hard and 25.4% on FrontierCode v2.0. DeepSeek-V4.1-Flash or Kimi K3 usually lead these rows, with Inkling second. On the PMPP-Hard GPU kernel optimization benchmark it solves 54.25 of 69 tasks, against 63 for GPT-6 Astra.

Vision is stronger: 83.8% on ScreenSpot Pro, 91.5% with a Python tool, 64.7% on PerceptionBench, and 1340 points on the DesignArena Website leaderboard, top among open models. In knowledge work it scores 70.5% on JobBench and 61.5% on the Vals Finance Index. Some results come from external evaluators (Artificial Analysis, Vals, Scale AI); the rest are self-reported.

In an autoresearch demo the model spent 4.5 hours and $426 of Tinker credit training a Qwen3.5-4B LoRA for time-series forecasting, ending with about 15% less error than the base model on unseen data.

Why it matters

For now it is only on Tinker for inference and fine-tuning. The company promised the weights under Apache 2.0 for the following week, while its model page now says the coming weeks. Once out, 17B active parameters keep inference relatively cheap, but 594B total needs serious memory. After Kimi K3.1 Pro, it is another strong open candidate (Kimi K3.1 Pro: Moonshot's open-weight model nears Opus).

Limitations by the company's own account: it is not very concise, it does not always pick the right tool in codebases with many skills, chat formatting is weaker, and it overreacts to mental health queries.