Can My Phone Run It?
Work out whether your phone can run a given local AI model. Uses the same RAM gate the TokForge app applies, plus peak memory measured on real handsets, with the error bars stated.
Pick the RAM your phone is sold with. Android reports slightly less than the marketing figure because the kernel reserves some, and the app rounds back up to the nearest standard tier, so a handset that reports 7.4 GB is treated as an 8 GB device. The iPhone or iPad option checks the Android catalogue against the tighter memory headroom on iOS. The iPhone app keeps its own catalogue, which also offers a 9B IQ2_M build from 8 GB and Gemma 4 E4B from 12 GB.
TokForge is free. Android is live on Google Play; iPhone and iPad are in public beta on TestFlight. No account, and it works with the network off once a model is downloaded.
Reading on a computer? See real measured speeds on the phone-speed leaderboard.
What this actually calculates
Two separate things, and they deserve different amounts of trust.
The compatibility verdict is exact. It is not a heuristic. TokForge snaps your reported RAM up to the nearest tier in the list 4, 6, 8, 12, 16, 24, 32, 48, 64 and then offers you every model whose floor is at or below that tier. This page runs the same rule against the same Android catalogue, so if it says the app will offer you a model on Android, the app will offer you that model.
The memory estimate is not exact, and every calculator that tells you otherwise is guessing with more decimal places. What we have instead of a guess is measurements.
The measurements the estimate is built on
These are peak resident memory readings submitted by real handsets through the benchmark in the app, matched against the download size the catalogue records for that exact build. These pairs come from runs submitted in July 2026; newer runs on the board are not included here.
| Build | Format | Device | Download | Peak resident | Ratio | Difference |
|---|---|---|---|---|---|---|
| Qwen3.5 0.8B | MNN | SM-G988U | 0.6 GB | 0.66 GB | 1.10x | +62 MB |
| Qwen3.5 0.8B | MNN | V2270A | 0.6 GB | 0.69 GB | 1.14x | +89 MB |
| Qwen3.5 2B | MNN | SM-A336E | 1.4 GB | 1.49 GB | 1.07x | +93 MB |
| Qwen3.5 4B (abliterated) | MNN | BVL-AN00 | 2.5 GB | 2.36 GB | 0.94x | -145 MB |
| Qwen3.5 4B (abliterated) | MNN | V2314A | 2.5 GB | 2.40 GB | 0.96x | -100 MB |
| Qwen3 4B (abliterated) | MNN | BVL-AN00 | 2.4 GB | 2.52 GB | 1.05x | +118 MB |
| Qwen3.5 9B (abliterated) | MNN | BVL-N49 | 6.1 GB | 4.32 GB | 0.71x | -1818 MB |
| Qwen3.5 9B (abliterated) | MNN | SM-S918B | 6.1 GB | 4.64 GB | 0.76x | -1491 MB |
Some further rows on the board report peak memory too small to be a real measurement of a model that size, for example 42 MB for a 9B model. Those are excluded here rather than averaged in, because a broken reading pulls an estimate down and nothing warns you it happened.
What the numbers say, and where they stop
For models under about 3 GB on disk, peak resident memory in these pairs landed close to the download size, a little above or a little below it. Treating the file size as the memory cost is close enough to be useful in that range.
Above that, the picture changes and our sample thins out badly. Both 9B measurements came in well below the file size, 4.3 and 4.6 GB against a 6.1 GB download, which is what memory mapping is supposed to do: pages that are not being touched do not have to be resident. Two samples is not a rule. The tool reports the spread rather than inventing a curve through two points.
This is also why the tool does not add a context or KV cache slider. On a desktop calculator that is the honest part of the arithmetic, because you can compute it from the layer count and the head dimensions. Here it would be a number we have not measured on a phone, sitting next to numbers we have, and looking equally confident. If we instrument it, it goes in.
Why phone memory is not desktop memory
Every calculator we could find is built for desktop GPUs and answers "does it fit in VRAM". Three things break that model on a phone.
You do not get all of it. The operating system, the launcher and everything you have not closed are in the same pool. Android documents a memory limiter that constrains how much a process may hold, with different ceilings for a visible app and a backgrounded one, and there is no runtime API to ask what your ceiling is. See Android's memory limiter documentation and its app memory overview.
Backgrounding costs you. A model that loads fine can be dropped when you switch to another app, and you pay the load time again on return. This is the single most common complaint about running large models on phones and it is not a bug.
iOS terminates rather than swaps. A process that crosses its limit is killed by jetsam. Apple documents the behaviour in Identifying high memory use with jetsam event reports. The practical effect is that iOS has less headroom than an Android phone with the same RAM figure on the box, which is why this tool warns on iOS earlier than it does on Android.
Common questions
Can my phone run a local LLM?
4 GB is the smallest tier TokForge serves, with Qwen3.5 0.8B and Llama 3.2 1B. Note that the app snaps your reported RAM up to the nearest tier rather than down, because Android reports less than the figure on the box once the kernel has taken its share, so a 3 GB handset is treated as the 4 GB tier and is offered those two models. Between 4 and 6 GB you are in small-model territory. At 8 GB the 4B class becomes usable, and 12 GB is where 8B and 9B models stop being an experiment.
How much RAM does a 7B model need on a phone?
More than the file size suggests, and the honest answer is that it depends on the build. On Android, TokForge offers 8B and 9B models from 12 GB, with one exception: Llama 3.1 8B appears from 8 GB, marked as large and slow, and loads only when you choose it. On the public leaderboard, two 9B runs recorded peak resident memory of 4.3 and 4.6 GB against a 6.1 GB download, because not all of a memory-mapped model stays resident at once. That is measured, not modelled.
Why is this different from a VRAM calculator?
A desktop VRAM calculator answers one question: does the model fit in the card. A phone has no card. Memory is shared with the operating system, the app is not free to use all of it, and the amount you are allowed shrinks the moment you switch away from the app. Android 17 introduced a per-app memory limiter with no runtime API to query it, and iOS will terminate a process that exceeds its jetsam limit rather than swap. Fitting is necessary and not sufficient.
Does the download size equal the memory it uses?
Close, but not reliably. In the device and build pairs we could match to a known download size, models under about 3 GB peaked close to their file size, a little above or a little below it, so treating the file size as the memory cost is roughly right there. The two 9B measurements both came in well under the file size. Two samples is not enough to make a rule from, so the tool says so instead of pretending otherwise.
Is this calculator accurate?
The compatibility verdict is exact, because it is the same gate the Android app applies: your RAM is snapped up to a standard tier and compared against the model's floor. The memory estimate is not exact, and the page shows the spread of the measurements it is built on rather than a single confident number. Where we have a measurement for the exact build you picked, the tool shows the measurement instead of the estimate.