DeepSeek V4.1 Flash on some tasks but remains behind it on several others.
But here's the experiment that really caught my attention.
Inkling 2 Small autonomously trained another AI model.
Thinking Machines gave it a challenge:
Take Qwen3.5-4B and improve its ability to forecast time series.
Things like weather patterns, electricity demand, and Wikipedia page views.
And then they let the agent work.
Here's what happened:
Inkling analyzed the forecasting problem.
It wrote its own evaluation code.
It discovered and corrected a bug in that evaluator.
It prepared training datasets.
It experimented with supervised fine-tuning (SFT).
It tried reinforcement learning (RL).
It compared checkpoints and selected the best-performing model.
The final result?
4 hours and 17 minutes of autonomous work
$426 spent through Tinker
Qwen3.5-4B successfully fine-tuned
More than 2× improvement in the internal forecasting score
Approximately 15% lower prediction error on previously unseen data
And that's a much more interesting result than simply winning another chatbot benchmark.
An AI agent researched, trained, evaluated, debugged, and improved another AI model.
Not perfectly. Not without limitations.
But with relatively little human intervention during the experiment.
And there's more.
Thinking Machines also tested Inkling 2 Small on GPU kernel optimization.
On PMPP-Hard, a benchmark involving 69 GPU programming challenges, the model solved an average of 54.25 tasks across four independent attempts.
That's close to DeepSeek V4.1 Flash at 55.25.
The model was also trained to handle extended coding tasks across different agent frameworks, with more emphasis on verifying its work rather than simply producing code that passes tests.
Now for availability.
Inkling 2 Small is currently available through Tinker Playground, including model sampling and fine-tuning.
The weights are still closed.
However, Thinking Machines plans to release the full weights under Apache 2.0, following external safety testing.
The company has indicated that this could happen within a week of the November 5 announcement, although the exact timing depends on those reviews.
And here's why I find this release particularly interesting.
For years, most AI benchmarks measured whether a model could answer difficult questions.
Now we're increasingly evaluating whether models can complete complex work.
Build a website.
Develop an application.
Optimize GPU code.
Run scientific experiments.
Or fine-tune another neural network.
The next major step for AI agents might not be generating better answers.
It might be independently improving the systems and tools we already use.
And watching a supposedly "small" model train another model in four hours is a pretty compelling demonstration of where things are heading.
594 billion parameters. Seventeen billion active. And apparently, Small is the new Large.
Official Inkling 2 Small announcement:
https://thinkingmachines.ai/news/inkling-2-small/
Try Inkling 2 Small on Tinker:
https://tinker.thinkingmachines.ai/
Thinking Machines Lab:
https://thinkingmachines.ai/
#Inkling2Small #ThinkingMachines #MiraMurati #OpenWeightAI #AIAgents #AutonomousResearch #LLM #GenerativeAI
YouTube / ABVX