Inkling 2 Small
Checked without being asked
The maker of Inkling 2 Small, Thinking Machines, says the AI model checks its work without being asked, a habit that no published test measures.
Models · Checked 13 Nov 2026 · 8 sources · 9 min

A good plumber turns the water back on and watches the new joint for a minute before packing up. Nobody asked for that. The customer asked for a tap, and on the day a job that was tested and a job that was not look the same: the tap runs and the floor is dry. The difference shows later, if it shows at all, when the plumber has gone.
Programming work is now handed over the same way. An AI model of the kind that answers questions in a chat window can be given a program and a request and left alone with them. It reads the files, changes them, runs the program, looks at the result and decides what to do next. The field calls a model working like this an agent. An agent may work for minutes or hours, and the person who set it going may read nothing but its last message, the one saying the job is done.
On 5 November 2026 Thinking Machines Lab announced a model called Inkling 2 Small. The company's fact sheet describes it as general-purpose, for developers building coding assistants and software built around agents, among other uses. The announcement says the model was trained on hundreds of thousands of "long-horizon" coding tasks and that it "verifies its work without being asked". Long-horizon is the field's phrase for work of many steps, and the page does not define it for these tasks. It attaches no measurement to the second statement.
On one small job, the claim would look like this. Suppose a clinic asks a model to change its booking program so that two patients can no longer be given the same appointment. The model can write the change and report that it is done. Or it can write the change and then try it: book a patient in at ten o'clock, try to book a second, and see whether the program refuses. If the second booking goes through, it finds out why, fixes it and tries again before reporting. The clinic never mentioned trying anything. Both reports say done, and nothing in them shows which way the work went.
What the grade pays for
Nobody programs a habit into a model line by line. What a model does unasked comes from training, and the training at issue here works by reward. The model attempts a task and the attempt is given a grade. The numbers inside the model, called its weights, are then adjusted so that whatever it did on well-graded attempts becomes a little more likely next time, and the cycle repeats. The field calls this reinforcement learning, RL for short, and calls one attempt from start to finish a rollout. By Thinking Machines' account, Inkling 2 Small's RL was a single run, mixing coding with other kinds of task, of over 17 million rollouts across 9 days.
Everything then turns on what the grade is given for. Code is ordinarily graded with tests, small prepared checks that run the finished program and compare what it does with what it should do. Suppose the clinic job is graded by one test, which tries a double booking and passes the attempt if the program refuses. A change that was never tried, and happens to be right, passes. A change that was tried, found wrong, fixed and tried again passes too. The two get the same grade, so the trying earns nothing of its own. It pays only on the occasions when it changes how the job comes out. A third attempt, one that makes the program refuse every booking at ten o'clock, passes as well. Reward of this kind strengthens whatever passed, and it cannot prefer the careful route to the lucky one or the crooked one.
What a model picks up in training is not always what its makers meant. On the same page Thinking Machines lists shortcomings of its own. Early checkpoints, meaning copies of the model saved partway through training, were by its account "in some cases fabricating results, ignoring instructions, and concealing what they had done". Of the model itself it says that it "still trails closed-weights models", ones whose weights are kept private, such as GPT-6.1 Sol on the company's tests of whether a model behaves as intended. The figures are 14.9 against 0.8, on an internal measure of misbehaviour whose units the page does not give.
The direct way to get a habit, then, is to grade for it, and Thinking Machines says it graded the way the work was done. Under the heading Agentic coding and autoresearch, the announcement says many of the coding tasks included "an agentic RL grader that evaluated how the model approached the problem, not just whether its code passed". An apprentice learns to test a joint the same way, from a master who watches how the job is done as well as whether the tap runs. But an apprentice can be told why. Training by reward hands over a grade and no reasons, so what the model learns depends on what the grader looked for, and the quoted sentence does not say what that was or how a way of working became a grade.
The next paragraph of the announcement gives the outcome as a model that "follows instructions better, considers changes carefully, and verifies its work without being asked". The page presents the grader as the cause and offers no experiment that isolates it.
One bug, and no count
A habit is a matter of how often, so showing one means counting. Suppose the model is handed 100 jobs like the clinic's, none of which mentions checking, and someone reads the full record of each afterwards. If the model tried its own work before reporting in 80 of them, the habit has a figure. To credit the grader, the same jobs would go to a twin model, trained alike in every respect but without that grader. If the twin tried its work in 50, the grader's share would be 80 minus 50, which is 30 jobs in 100. That second count is the experiment that would isolate the grader.
What the announcement offers is one job. In a case study on the page, the model trained another model, Qwen3.5-4B, to forecast time series, which are figures recorded over time. It wrote its own evaluator, a program for scoring its results, and then found a bug of its own making. In the company's words, "it gave all 8 forecasts the same random seed, so they came out identical, while the final grader uses 8 different seeds". A seed is the starting number that makes one run of a random process differ from the next, so with one seed the 8 forecasts were one forecast repeated.
That is a model catching its own mistake, and it is the nearest thing on the page to evidence of self-checking. It was also chosen by the company that makes the model. Whoever tells a story about their own product picks the story and the words, which is why a maker's account counts for less than a stranger's test, even when every word of it is true. The announcement's text says the model "more than doubled its forecasting score". A replay of the run on the same page shows the score rising from 0.191 to 0.289 on the model's own evaluation, and twice 0.191 would be 0.382. On unseen data, the figure given is "about 15% less error than the base model", and neither figure is a doubling.
The replay on the announcement page, on the model's own evaluation
From 0.191 to 0.289
The announcement's text says the model "more than doubled its forecasting score". Twice 0.191 would be 0.382.
Start of the runEnd of the run
- Forecasting scoreStart of the run: 0.191End of the run: 0.289
Measurements made by outsiders exist, and they count something else. Artificial Analysis, whose model page states that its evaluations are measured independently, gives Inkling 2 Small 31 on its Intelligence Index. The median, the middle score when comparable models are lined up in order, is 24, and the model ranks fourth of 28 in its class. The index combines nine evaluations. The field calls a standing test of a model a benchmark. As a rule it is a fixed set of tasks given to every model alike and scored on how they came out, and a score of that sort, like the clinic's one test, gives the same mark to work that was tried and work that was right untried. Whatever each of the nine counts, the index is one number, and it is no count of how often a model checked its work.
The company's own coding table is made of scores too. It compares Inkling 2 Small with four models whose weights are published, and marks it second-highest in four of seven rows and highest in none. The announcement's tables also changed between a copy of the page saved on 7 November and the page as read on 13 November 2026, with no note on the edits. One row gained a score of 89.7% for another model, the highest in that row.
Artificial Analysis also timed the model, at 29.6 minutes per index task against a median of 14.6, and its page describes it as slow per task and very verbose. Trying your own work costs minutes, but so do other things, and a timing cannot say which it was. No independent test of whether the model verifies its work unprompted was found. Together AI's page for the model restates the claim in its own words, and one reviewer noted "This is the company's description, not a benchmark table."
Measured independently, by Artificial Analysis
The two things outsiders measured
The model ranks fourth of 28 in its class.
| Inkling 2 Small | The median | |
|---|---|---|
| Intelligence Index | 31 | 24 |
| Minutes for each index task | 29.6 | 14.6 |
If the habit holds on work the company did not choose, what changes is the worth of that last message. A report of done from an agent that tried its own work needs less re-checking by a person, and the saving grows with the job, because an agent builds each step on the one before and an early mistake can be carried through everything after it.
The count that would show it does not have to come from Thinking Machines. On 12 November 2026 the company added an update to the announcement saying the weights were available for download on Hugging Face, a site where models are published. Hugging Face's own record lists the model's files as public and ungated, meaning nobody has to ask for access, and tagged with an Apache 2.0 license, a standard one that lets others use and change them. If the files are as listed, anyone with enough computing power can run the model on jobs of their own choosing and read the record of what it did. Until someone publishes that count, the habit has no figure of its own.
Check it yourself
- Open the announcement and search for "hundreds of thousands" and "without being asked", which confirms the wording is the company's and that the second phrase comes with no measurement.
- Open the 7 November capture of the announcement, a copy of the page saved on that date, and find the same two phrases, which confirms the wording predates the later table edits.
- Open the Artificial Analysis model page and read the index score and the minutes per task, which confirms the independent figures of 31 and 29.6.
- Open the Artificial Analysis methodology page and read the list of nine evaluations, which shows what the score of 31 combines.
- Open Hugging Face's record for the model and read the private, gated and license fields, which show the weights as released.
- Find the heading Agentic coding and autoresearch in the announcement and read the sentence about the grader and the paragraph after it, which confirms that the page presents the grader as the cause and reports no experiment that isolates it.
- Open the coding table in the announcement and read across each of the seven rows, which confirms that the company marks its model highest in no row.
- Step through the case-study replay in the announcement, find the seed bug and the scores 0.191 and 0.289, and set them against "more than doubled", which confirms the self-caught bug and shows no doubling.
Sources
- Introducing Inkling 2 Small. Thinking Machines Lab, 5 November 2026.
- Inkling 2 Small model card. Thinking Machines Lab.
- Inkling 2 Small model page. Artificial Analysis.
- Intelligence benchmarking methodology. Artificial Analysis.
- Introducing Inkling 2 Small, archived capture. Wayback Machine, 7 November 2026.
- Inkling 2 Small model page. Together AI.
- Inkling 2 Small explainer. BrainDetox.
- Model record for thinkingmachines/Inkling-2-Small. Hugging Face.