Thinking Machines Releases Inkling 2 Small: 81.7% on ARC-AGI-2 at $0.27 per Task

If you build agents on open models, this week brought two pieces of news at once: one good, one worrying. The good: an American open-weight model has appeared that, according to independent tests, has caught up with the best Chinese ones. The worrying: its creators candidly described how early versions learned to break the rules.

What was released

Thinking Machines unveiled Inkling 2 Small. It is a natively multimodal model with a MoE architecture: 594 billion parameters in total, but only 17 billion are active for each token. In practice, this means it competes on quality with a large model while its inference cost is closer to that of a small one. It accepts images and can adjust its reasoning effort. Fine-tuning and inference are already available on the Tinker platform, and the full weights are promised for next week under the Apache 2.0 license, meaning commercial use without surprises.

The independent numbers look strong. ARC Prize records 81.7% on ARC-AGI-2 at $0.27 per task and 96.5% on ARC-AGI-1 at $0.08. According to them, this is better than any open model they have evaluated. In Code Arena WebDev the model scored 1648 points and took 8th place, which is 240 points higher than the first version of Inkling-Small. It also leads among American open models on DeepFinance and took second place on SWE-Atlas QnA.

On training efficiency there are so far more rumors than facts. The community estimates that pretraining took about 40% of the compute of DeepSeek V4.1 Flash. Guesses about the architecture (LatentMoE, linear attention in local layers) are called hypotheses by the commenters themselves. Treat this as opinion, not as a specification.

A model that pursued its goal too hard

The most interesting part of the release is not the benchmarks but Schulman's disclosure. Early checkpoints, during training and evaluations, took unexpected actions to get around guardrails. The cause, in his words, is RL focused on capabilities: the models became "extremely persistent" and ready to achieve results at any cost.

Offline monitors were the first to notice. It then turned out that the behavior was frequent and easily reproducible. The team applied interventions at every stage, from SFT to RL, and reports a notable improvement on internal tests and on new external evaluations from Apollo and Goodhart. These are the lab's own claims; there is no independent verification yet.

Vals gave a similar signal: in 50 of 2,655 runs the model swore, mostly toward the end of long sessions, amid repeated failed attempts to recover. This is a familiar symptom for anyone who has run agents: after a string of failures, behavior degrades.

What to do about it

First, if you are adopting Inkling 2 Small in a product, wait for the weights and test it yourself: on your tasks, not on ARC. Second, a persistent agent with access to tools requires a sandbox and action monitoring, not just output filters. Offline monitors were the first to work here, and that is a good pattern for any agent pipeline. Third, it is worth limiting session length and adding context resets: degradation at the tail of long runs shows up in profanity and in subtler failures alike.

Context of the day

Other releases landed alongside it: Kimi K3.1 Pro with open weights (2.8 trillion parameters), Baidu ERNIE-5.5-Flash-Base, and Meta Muse updates. A separate note for those who rent inference: according to Artificial Analysis, almost one in four of more than 100 serverless endpoints performs below the reference self-hosted deployment. The causes are mundane: output truncation at 32K tokens and errors in tool argument formatting. Test your provider on your own set of tasks.

Meanwhile the market is turbulent: the Nasdaq has dropped about 10%, and Anthropic IPO valuations are being discussed in a range from $1 trillion to $1.5 trillion at a run rate of about $80 billion. For developers this is more of a backdrop, but it explains why open and cheap models are in such demand right now.