Efiwe 1.5 is 1.4 MB in download size. The smallest of the ten external language models you can now download on Efiwe is 105 MB, and the largest is 1.1 GB. That size gap explains almost everything about how the two kinds of model share the work.
On October 12, 2026, we added ten external small language models (SLMs) that run on your phone: SmolLM2 135M, LFM2 350M, Gemma 3 270M compact, SmolLM2 360M, Gemma 3 270M, LFM2 700M, Qwen 2.5 0.5B, LFM2 1.2B, Gemma 3 1B and Qwen 2.5 1.5B. They write fuller, friendlier explanations. Efiwe 1.5 still grades your code offline and whenever an on-device model is selected, and it still works out what is wrong with it, across all eight courses: HTML, CSS, JavaScript, Tailwind, Python, PHP, SQL and Prompt Engineering. This article explains why, and how to pick the right setup for your phone.
Two different jobs: checking and explaining
Efiwe 1.5 is a checker. The ten external SLMs are explainers. They are not competing for the same job.
Efiwe 1.5 was built for one thing: coding education. It runs entirely on your phone, needs no internet, and compares your code, and what it produces, with the correct solution. It knows the rules of each language it teaches. It can tell a missing colon in Python from a wrong indentation level, a missing semicolon in PHP from a wrong operator, and “= NULL” from “IS NULL” in SQL. Give it the same code twice and it gives the same answer twice.
The ten external SLMs are general-purpose language models from Hugging Face, Liquid AI, Google and Alibaba’s Qwen team. They were trained to write fluent text on almost any subject, not to grade code. That makes them good at turning a dry finding like Missing “alt” attribute in <img> tag into a sentence a beginner enjoys reading. It also means that, like every language model, they can sound confident while being wrong.
At a glance
Every model below runs on your phone, works offline once downloaded, and keeps your code on the device. Sizes are download sizes.
| Model | Made by | Size | Times larger than Efiwe 1.5 | What it does on Efiwe |
|---|---|---|---|---|
| Efiwe 1.5 | Efiwe | 1.4 MB | 1x | Grades every submission offline or with an on-device model, backs up Efiwe 2.0 online, finds every issue, writes hints, answers when any other model cannot |
| SmolLM2 135M | Hugging Face | 105 MB | 75x | Explains Efiwe 1.5’s findings |
| LFM2 350M | Liquid AI | 229 MB | 164x | Explains Efiwe 1.5’s findings |
| Gemma 3 270M compact | 253 MB | 181x | Explains Efiwe 1.5’s findings | |
| SmolLM2 360M | Hugging Face | 271 MB | 194x | Explains Efiwe 1.5’s findings |
| Gemma 3 270M | 292 MB | 209x | Explains Efiwe 1.5’s findings | |
| LFM2 700M | Liquid AI | 469 MB | 335x | Explains Efiwe 1.5’s findings |
| Qwen 2.5 0.5B | Qwen team, Alibaba | 491 MB | 351x | Explains Efiwe 1.5’s findings |
| LFM2 1.2B | Liquid AI | 731 MB | 522x | Explains Efiwe 1.5’s findings |
| Gemma 3 1B | 806 MB | 576x | Explains Efiwe 1.5’s findings | |
| Qwen 2.5 1.5B | Qwen team, Alibaba | 1.1 GB | about 790x | Explains Efiwe 1.5’s findings |
The number in each SLM’s name (e.g. 135M, 1.5B) is its parameter count: roughly how many adjustable values the model holds. More parameters usually means richer writing, more memory and slower answers. None of the ten external SLMs decides whether you pass.
Why 1.4 MB matters
Efiwe was built for learners whose data is expensive and whose phones are modest. A model that costs a few seconds of data and a sliver of memory can be on every phone. A model that costs half a month’s data cannot.
- Data. The ITU reports that the average mobile broadband subscription in a low-income country carries 2.2 GB per month. Qwen 2.5 1.5B alone is half of that. Efiwe 1.5 comes with the platform and costs nothing extra to download.
- Memory. Larger models need more memory while they run. Some phones, especially iPhones, close the page when a model is too big. Efiwe remembers this and steers you to a lighter setting next time. Efiwe 1.5 is small enough that this never happens.
- Speed. Efiwe 1.5 answers instantly. The ten external SLMs take a few seconds, longer on older phones. If a model takes too long, Efiwe 1.5 answers instead, so you are never left waiting.
- Battery. Generating text word by word is hard work for a phone’s processor. Checking code against a solution is light work.
That is why Efiwe 1.5 is always there, and the ten external SLMs are an optional upgrade you choose to download, ideally once over Wi-Fi.
Why Efiwe 1.5 does the grading and the diagnosis
A grade has to be right, and it has to be the same every time. Efiwe 1.5 is built to be both. A general-purpose SLM is built for neither.
It does not invent. Efiwe 1.5 can only report differences that are actually there between your code and the solution. An SLM generates the most likely next words, so it can describe a bug that does not exist, or miss one that does. On a grade, either mistake costs a learner points and lives, or passes them on something they have not learned.
It is consistent. The same code always gets the same verdict, online or offline, on any phone. That matters for points, the leaderboard, badges and certificates, which have to mean the same thing for every learner.
It knows each language, and is fair about what does not matter. It ignores differences that do not change what the code means, and catches the mistakes beginners actually make:
| Course | What Efiwe 1.5 ignores | What it catches |
|---|---|---|
| HTML and Tailwind | Spacing, attribute order, class order, curly “smart” quotes, an extra matching closing tag | Missing or extra tags and attributes, wrong attribute values, tag name typos |
| CSS | Declaration order inside a rule, spacing, quotes, letter case (rule order still counts, because later rules override earlier ones) | Unclosed braces, invalid syntax, wrong values for a selector |
| JavaScript | Indentation, line breaks, optional semicolons, trailing commas, quote style, a + b written as b + a | Wrong function called, function name typos, wrong let, const or var, missing statements or blocks, mistakes inside if/else and loops |
| Python | Indent width (2 spaces, 4 spaces or a tab), blank lines, quote style, trailing commas (block structure still counts) | Missing colons, wrong indentation level, wrong operators, numbers or strings, missing or extra lines |
| PHP | Layout, letter case of keywords and function names, equivalent quotes, the closing ?> tag | Missing semicolons, wrong operators; switch cases, match arms and independent assignments are accepted in another order when that provably changes nothing |
| SQL | Layout, letter case of keywords and names, the final semicolon | Missing semicolons, wrong operators, “= NULL” instead of “IS NULL”; INSERT rows and IN-list values are accepted in any order |
| Prompt Engineering | Wording, and the language you write in | Missing required details, techniques the challenge teaches but you did not use, vague wording, run-on prompts, telling the AI not to do the task, asking for something then taking it back |
It scores against a clear bar. Efiwe 1.5 runs your code and the solution and compares what each produces. If they match, you pass, so there is more than one right way to solve a challenge.
It prioritises. Instead of listing everything at once, it ranks issues by how serious they are and how often you have encountered them, and shows the most important few first.
This is the knowledge we do not hand over to a general model. We hand over its results.
What happens when you tap each button
With an on-device model selected, Efiwe 1.5 always does the checking. An external SLM when selected only ever writes the words.
| Button | Who decides the facts | Who writes the answer | If the SLM cannot answer |
|---|---|---|---|
| Submit | Efiwe 1.5 | Efiwe 1.5 | Not used. Submit goes straight to Efiwe 1.5 |
| Diagnose Issue | Efiwe 1.5 | Your chosen SLM | Efiwe 1.5 shows its own list of issues |
| Show Hint | Efiwe 1.5 | Your chosen SLM | Efiwe 1.5 shows its own hint |
| Learn More | Not about your code | Your chosen SLM explains the challenge’s topic | Efiwe falls back to its standard reference content |
Submit never waits on an SLM. Marking stays instant, private and identical whether you are online or offline, and it never uses up the free plan’s five-minute cloud AI cooldown.
With Efiwe 2.0 (cloud) selected and online, Submit is graded by the cloud model, and Efiwe 1.5 takes over if it does not answer. In Prompt Engineering, Efiwe 1.5 still sets the limits: Efiwe 2.0’s score can be at most 25 points above it, a missing required detail still fails the attempt, and a prompt that tries to talk Efiwe 1.5 into a higher score is caught.
Learn More is the one place an external SLM writes from its own general knowledge, because it explains a concept rather than your code. It is background reading, never a verdict.
How the handoff works
The ten external small models make things up, so we never ask them to guess. When you tap Show Hint or Diagnose Issue with an external SLM selected:
- Efiwe 1.5 checks your code against the solution and lists the facts: what is missing, what is misspelled, what is extra.
- Efiwe hands the external SLM those facts, together with your code and the solution. The SLM’s only job is to explain them in plain, friendly words.
- Efiwe checks the answer from the external SLM. If it is empty, cut off, repeats the question, or misses the facts, it is dropped.
- If the answer is dropped, or the model is too slow or cannot load, Efiwe 1.5 answers instead.
- The answer is labelled with the model that wrote it, so you always know where it came from (e.g. Generated by Efiwe 1.5)
This extra check does not count against you. Efiwe 1.5’s look at your code for the external SLM is not recorded as a second mistake in your error history, which Efiwe uses to choose your hints.
If you use Efiwe 2.0 (cloud) and lose your connection, Efiwe can hand the question to an external small model you have downloaded (and selected) before falling back to Efiwe 1.5. The same happens if Efiwe 2.0 fails. It does not happen when you hit the free plan’s five-minute cooldown, so that limit works exactly as before.
Which should you choose?
Pick from the AI model dropdown under any challenge. Your choice is saved to your account, so it follows you across courses and shared devices.
| If you… | Choose |
|---|---|
| Are on a metered data bundle or an older phone | Efiwe 1.5 |
| Want the fastest possible feedback | Efiwe 1.5 |
| Want fuller explanations offline and have Wi-Fi to download once | SmolLM2 360M or LFM2 350M to start |
| Have a newer phone and want the richest offline explanations | Qwen 2.5 1.5B or Gemma 3 1B |
| Are online and want the most detailed help | Efiwe 2.0 (cloud) |
Efiwe 1.5 and the ten external downloaded small models do not use the free plan’s five-minute cooldown, so free learners get unlimited instant feedback with either. Downloads resume where they stopped if your connection drops, and you can remove a model at any time to free up space.
Recap
Efiwe 1.5 is the judge. The ten external SLMs are the storytellers. Efiwe 1.5 is 1.4 MB, built for coding education, and decides what is right and what is wrong whenever you learn offline or with an on-device model. The ten external SLMs are 75 to roughly 790 times larger than Efiwe 1.5, and their job is to explain Efiwe 1.5’s findings in warmer words. Keeping those jobs apart is how Efiwe gives you richer explanations without ever letting a guess decide your grade.
Open any challenge, pick a model, and try a lesson with your data switched off. Tell us how it goes at product@efiwe.com.
Credits: SmolLM2 by Hugging Face and Qwen 2.5 by the Qwen team (Apache 2.0), Gemma 3 by Google (Gemma Terms of Use), and LFM2 by Liquid AI (LFM Open License 1.0). The models run in the browser through wllama, built on llama.cpp (MIT). Data figure: ITU, Measuring digital development: Facts and Figures 2025.
or