Aleph Alpha's Kolibri lets users dial reasoning effort from none to high
Aleph Alpha's recently released Kolibri model lets users choose how long it reasons, with four settings: none, low, medium or high. The company says its training data had no effort labels, so it annotated every reasoning trace after the fact by comparing its length with similar traces.
- Kolibri is a hybrid model with none, low, medium and high reasoning effort, and high is the default.
- Effort is set by one sentence in the system prompt, not a special token.
- Training traces were labelled post-hoc by length relative to groups of the same dataset and judged difficulty.
- On 23 of 24 benchmarks the SFT checkpoint reasons longer at medium than low and longer still at high.
- At low, the SFT checkpoint keeps 96% of its high score with 35% of the reasoning length, according to Aleph Alpha.
A single hybrid model with a dial
According to Aleph Alpha, Kolibri is a single hybrid model that either answers directly or reasons at low, medium or high effort, with high as the default. The company places it alongside Qwen3, Llama-Nemotron and GPT-OSS, which follow a similar hybrid approach.
The reasoning_effort parameter is implemented as one sentence in the system prompt rather than a special token. Aleph Alpha says it chose a sentence because the model can follow it before it has seen a single effort label. The sentences range from "Reasoning is disabled" for none to a high setting that asks the model to consider plausible alternatives and prioritize correctness and clarity.
The motivation is latency. Longer reasoning is known to raise accuracy on hard problems, but the post notes that a reasoning model asked what 2 + 3 is may produce a page of thinking before answering 5. Aleph Alpha says it wanted behaviour that adapts to prompt difficulty on its own, plus a dial for users who value speed over accuracy.
Labelling effort after the fact
The supervised fine-tuning (SFT) traces did not record how much effort went into them, so the team labelled each one by measuring its reasoning length. It first changed the unit of measurement. Kolibri Origin, the internal predecessor, counted characters, but one reasoning token carries between 2.43 and 5.69 characters, with math at the low end. A 2,000-character threshold therefore meant 822 tokens in one dataset and 351 in another. Kolibri now counts tokens.
The team also defined an exchange as all reasoning the assistant does between two turns a person writes. A tool loop with twenty calls counts as one exchange.
A simple split into the lowest 30%, middle 30% to 70% and top 30% of lengths proved poor, according to the post, because length depends mostly on domain. The median chat trace has 310 tokens and the median math trace 8.2K. With pooled thresholds, 49% of chat became low and 83% of math became high, so the label mostly revealed the source dataset.
Setting thresholds per dataset helped, since median traces across 74 datasets run from 27 to 90K tokens. But spread remained within datasets. In one math dataset, the middle 95% of traces runs from 430 to 79K tokens.
Accounting for difficulty
To capture difficulty, Aleph Alpha used GLM-5.2 as an LLM judge on 100,000 conversations, grading them easy, medium or hard. It distilled those verdicts into a 2B student model, similar to DEITA. The student grades each conversation four times, the grades are averaged and rounded toward the harder one. The company says a hard request with too little reasoning costs more than an easy one with too much. The student agrees with GLM-5.2 84.8% of the time.
Difficulty explains much of the spread. In the math dataset, medians by grade are 592, 2.6K and 18K tokens. Across datasets, hard prompts reason up to 31.3 times longer than easy ones. The factor is 21 in code and chat and 5.5 in science, while tool use and safety barely react.
Thresholds are fitted per group, meaning one dataset at one difficulty grade. 176 groups in 69 datasets reach the required 200 exchanges and each gets its own 30th and 70th percentile. Smaller groups and ungraded conversations fall back to dataset thresholds.
To avoid a hard cut that a model could learn as the place to stop thinking, labels are drawn probabilistically. A trace's length is placed on a 0 to 1 interval by interpolating in log space through the group's two percentiles, and the label probability falls off as a Gaussian with distance from each bucket's centre.
Results and what it means
Aleph Alpha reports that on 23 of 24 benchmarks its SFT checkpoint reasons longer at medium than at low and longer still at high. At low, it keeps 96% of its high score while reasoning 35% as long. The released model keeps 98% with 76% of the completion tokens, according to the company.
For users, the practical point is that effort trades quality against latency without requiring them to judge how hard their own prompt is, since the model works out how much reasoning a request needs. These figures come from Aleph Alpha's own evaluation.
Aleph Alpha: Reasoning Effort That Adapts to the Request — Aleph Alpha
