Can a Laptop With No Internet Teach El Salvador's 4th Grade?
Since 2021, El Salvador has handed out between 1.2 and 1.6 million laptops and tablets to public-school students, and the students keep them at home. That's the part that caught my attention. The device is already in the house. What usually isn't there is internet: only 37% of Salvadorans have internet at home, and in rural areas it's 13%.
So here's the question this series is about:
Can a small AI model, running offline on the device a student already has, tutor the official school program?
I'm calling the project Narrow Intelligence. It's an open-source, non-profit experiment, and El Salvador's 4th grade is the worked example. The code will be published.
Why offline
There is already an AI tutor in Salvadoran public schools: the government announced a cloud tutor with xAI's Grok in December 2025. A cloud tutor needs a connection, and a school has one more often than a home does. So I see an offline tutor as the home complement: the same student, the same device, at the kitchen table at 8 p.m., with no signal.
Offline also changes the privacy picture. If the model runs on the device, the child's questions never leave it.
Not one model that knows everything
The obvious design is "take a small model and teach it the whole 4th-grade program". I think that's the wrong design. A model with 2 billion parameters can't hold a country's geography, history and math reliably, and when you train facts into a small model, it tends to invent more, not less.
So the tutor is three parts, each doing the job it's good at:
| Part | Holds | How |
|---|---|---|
| A small model (1–4B parameters, Apache 2.0) | the behaviour: explain, ask, never hand over the answer | a prompt in this baseline; LoRA training after |
| An on-device library | the knowledge: El Salvador's facts | searched before every answer |
| Checkers | what's right | SymPy for math, answer keys |
The model is the teacher's manner. The library is the textbook. The checkers are the answer key at the back of the book.
The pilot: two subjects, two different problems
I picked two 4th-grade subjects because they test two different things:
- Math tests behaviour. Any decent model can add 3/10 and 4/10. The hard part is not telling a 9-year-old the answer, and guiding them to it instead.
- Ciudadanía y Valores (civics and values) tests knowledge a model can't have: El Salvador's departments, its population, its rivers, the Nahua and Lenca peoples. A model trained mostly on English internet text has seen very little of that.
The map of what to teach is the official 2026 study programs: 98 numbered learning goals in 4th-grade math and 28 in Ciudadanía y Valores. I refer to them by number only (like "4° 8.5"). The official books are "all rights reserved", so I don't train on them or copy their text.
Test sets that grade themselves
Before training anything, I needed to measure the untrained models. That meant test sets, and test sets are where projects like this usually cheat without noticing. So every item had to pass a machine check, not my judgement.
Math: 325 problems we wrote ourselves. For each goal, a large model (DeepSeek V4 Flash) drafts problems in Salvadoran Spanish, each with a SymPy expression and an answer. A second call solves the same problems blind. A problem is kept only if all three agree: the SymPy result, the drafted answer and the blind answer. If a problem repeats 8 or more consecutive words of the official goal, it's dropped. A typical one:
Para pintar una manta de la escuela, pintas 3/10 de la manta de azul y 4/10 de la manta de blanco. ¿Qué fracción de la manta pintas en total?
Ciudadanía y Valores: 138 questions from a 50-article library. The library is 50 Spanish Wikipedia articles about El Salvador, Central America and its peoples, each pinned to a revision. For each goal, a search picks passages and the large model writes questions about them. An answerable question is kept only if its evidence appears word for word in the passage and a second check finds it clear and suitable for a 9-year-old. It also writes questions the library can't answer, kept only if a judge reading the library's best matches finds no answer. Those are the trap: the right reply is "no está en mi material".
Whole units are held out. The test set isn't a random sample of the same kinds of problems. Entire units are kept apart (math units 4, 6 and 7; civics units 2 and 4), so a test problem is a new kind of problem, not a near-copy of something the model practised on.
What I measured
Four untrained models, all small enough for a school device: Qwen3.5-0.8B, Qwen3.5-2B, Qwen3.5-4B and Gemma 4 E2B. Four questions:
- Can it solve the math?
- When it plays tutor, does it give away the answer?
- Without the library, does it know El Salvador, or does it invent it?
- With the right page from the library, does it answer correctly, and say "no está en mi material" when the page doesn't have it?
The short version of the results:
- Every model gives away the answer when it plays tutor: 49–100% of replies state the final answer before the student has tried. One line in the prompt fixes it for the 4B and the E2B, but not for the 2B. That's part 2.
- Without the library, the models invent El Salvador: the Qwen 2B and 4B invent the answer to 77–94% of Salvadoran questions. With the right page, 2B, 4B and E2B get 95–100% right. That's part 3.
- Speed: on a CPU with 4 threads, the 2B writes 39 tokens per second, and a tutor reply takes 10–12 seconds for the 2B, 4B and E2B. That's an Apple M1 Max, so it's an upper bound: a school laptop will be slower, and I haven't measured one yet.
What this isn't (yet)
I want to be precise about the claims, because education deserves it:
- These are untrained baselines. Nothing has been trained yet.
- No child has used it. These are test sets, not classrooms.
- Nobody has published learning results for a fully offline AI tutor. A real pilot, with a control group, would be a first. That's a long way off.
- The sets are small: math scores carry about ±12 points of uncertainty, civics ±15–20. Read small differences as ties.
There's also a third piece I'm not touching yet: a Náhuat taster waits for the language community's consent.
Next: part 2, small models give away the answer, and why that means the 2B has to be trained.