Back home

Ampersand Tutor

x 0.86
Visit source

Intro

Ampersand Tutor (written &Tutor) is a tutoring model released by the AI lab humans& on October 30, 2026. Instead of being an all-purpose assistant, it is trained specifically to teach — to help a learner actually understand a problem rather than simply hand over the answer.

Its distinctive training trick: humans& first built Persimmon, a large "user model" that simulates how real people behave and learn, and then trained &Tutor with reinforcement learning to teach Persimmon. The reward is how much the simulated student learns. &Tutor is available through the humans& agent API as ampersand-tutor-v0.1.

Explain

Key facts

  • Developer: humans& (an AI research startup focused on human-centric models).
  • Base model: built on GLM-5.3, a mixture-of-experts transformer (78 layers, 256 routed experts + 1 shared, 8 active per token), served in NVFP4.
  • Modality: text in, text out; evaluated with a 131K-token context.
  • Training: RL in simulated tutoring conversations. Persimmon plays the learner; a judge measures the learner's learning gain. There are no hand-written rubrics for "scaffolding" or "don't give away the answer" — those behaviours emerge from the objective.
  • Results (vendor-reported): on the TutorMoments benchmark, &Tutor scores 0.754 with a plain prompt vs 0.500 for its GLM-5.3 starting point, ahead of Claude Sonnet 5.5 (0.667) and GPT-6 Astra (0.634). With an evaluation-aware prompt it scores 0.919, just behind GPT-6 Astra (0.926), and has the best "appropriate rigor" score.
  • Ablation: training against GLM-5.3 as the simulated learner instead of Persimmon performed significantly worse, which humans& uses to argue that a better model of people yields a better teacher.

Caveats

All scores come from simulated students and LLM judges, not measured outcomes with real learners. humans& itself notes the results may not generalize to real students, other subjects or long-term learning.

Usage

How to try it

  1. Read the launch post: Introducing &Tutor and the model card.
  2. Request access to the humans& agent API, then call the model ID ampersand-tutor-v0.1.
  3. Send the conversation history plus tutoring instructions (subject, learner level, goal). The model returns the next tutor reply.
  4. Good fits: homework-help apps, practice tutors and teacher-assist tools where you want hints and questions instead of direct answers.
  5. Keep a human in the loop for anything consequential (grading, placement), as the model card advises.

Related: Persimmon (the simulated-learner model) has its own playground at persimmon.humansand.ai.