Gemma 3n E4B on a phone: RAM requirements
8GB-class phone runs it, razor-tight · 6GB takes Q2_K · no LiteRT build ships
Gemma 3n E4B's recommended Q4_K_M build needs ~4.7 GiB of free RAM at 4K context — 3.9 GiB of measured weights plus a 0.27 GiB KV allowance and a 0.50 GiB runtime reserve. Count usable RAM at 60% of the sticker and the door for that build is an 8GB-class phone, razor-tight — 0.1 GiB to spare. The reason a "4B-class" model files this heavy: MatFormer nesting and per-layer embeddings bring the total parameter count to 7.8B (4.4B effective), the same naming trap as Gemma 4 E4B. 6GB phones take the Q2_K build tight; 4GB is out; and unlike its siblings this model has no LiteRT build — litert-community does not ship one (negative existence, probed 2026-09-16), so GGUF via PocketPal or llama.cpp is the door. Phones have no swap: past the line the app dies, it does not slow down.
bartowski Q4_K_M GGUF (3.9 GiB), keep context at 4K.Gemma 3n E4B is Google's mobile-first entry in the Gemma 3n family (released 2025-06): a MatFormer model that nests an E2B configuration inside the E4B weights, 35 layers of GQA attention with 32K context. The nesting plus per-layer embeddings is why the measured Q4_K_M file reads 3.9 GiB against the "4B" name — the total parameter count is 7.8B. The official google repo is license-gated (metadata pinned, files 403 with our token), so the official-caliber number below comes from unsloth's byte-mirror of the official weights; GGUF ladders are from bartowski and unsloth at pinned revisions.
Data verified · sources pinned in Where every number comes from
How much RAM at each build? (weights + 4K KV allowance + reserve)
| Build | Weights | Total needed | Smallest phone tier | Note |
|---|---|---|---|---|
| Q2_K | 2.6 GiB | 3.3 GiB | 6 GB tier | the 6GB-class door — the only tracked way in at 6GB measured |
| IQ3_XS | 2.9 GiB | 3.7 GiB | 8 GB tier | the 8GB-class step-up measured |
| Q4_K_M — recommended | 3.9 GiB | 4.7 GiB | 8 GB tier | community default quality/size point — recommended measured |
| Q5_K_M | 4.6 GiB | 5.4 GiB | 12 GB tier | the 12GB-class ceiling measured |
| Q6_K | 5.3 GiB | 6.1 GiB | 12 GB tier | the 16GB-class ceiling measured |
| Q8_0 | 6.8 GiB | 7.6 GiB | 16 GB tier | the 24GB-class quality ceiling measured |
| LiteRT (.litertlm) | — | — | no build shipped | negative existence: litert-community ships no gemma-3n build (author search, 12 hits probed 2026-09-16) — the LiteRT door is stated as absent; GGUF is this page's door reported |
Caliber: this page counts in GiB throughout — the site-wide internal caliber — and applies the 60% usable rule to the sticker number as-is (4 → 2.4 GiB usable). Tier names keep the marketing GB spelling; every computed cell above derives from the pinned file bytes in provenance.
Is your phone enough? (find your phone class: one row tells you)
| Phone class | RAM | Usable | Biggest build that fits (ceiling) | Verdict |
|---|---|---|---|---|
| 4GB budget Android | 4 GiB | 2.4 GiB | none | NO · smallest tracked build is 3.3 GiB total |
| 6GB class | 6 GiB | 3.6 GiB | Q2_K (3.3 GiB) | TIGHT · razor, 0.3 GiB spare |
| 8GB — iPhone 15/16 Pro tier | 8 GiB | 4.8 GiB | Q4_K_M (4.7 GiB) | TIGHT · razor at Q4_K_M, 0.1 GiB spare |
| 12GB — Galaxy Ultra / iPhone 17 Pro tier | 12 GiB | 7.2 GiB | Q6_K (6.1 GiB) | TIGHT · 1.1 GiB spare at Q6_K |
| 16GB — Pixel Pro tier | 16 GiB | 9.6 GiB | Q8_0 (7.6 GiB) | TIGHT · 2.0 GiB spare at Q8_0 |
| 24GB — OnePlus flagship tier | 24 GiB | 14.4 GiB | Q8_0 | OK · 6.8 GiB spare |
Each phone class has exactly one ceiling — the biggest tracked build that fits. KV cache is counted at 4K context with a runtime reserve — phones have no swap, so there is no soft landing past the line. iPhone 15/16 Pro tiers sit at 8GB; the iPhone 17 Pro generation moved the Pro tier to 12GB (Sept 2025), and Android flagships carry the 12–24GiB end.
What phone should you buy for it? (the 8GB razor logic)
| The logic | 8GB is the cheapest tier that runs the recommended Q4_K_M at all (4.7 GiB total vs 4.8 GiB usable) — and it is a razor. At 6GB, Q2_K is the only tracked door. 12GB steps up to Q6_K tight; 16GB takes Q8_0 but stays amber; 24GB is the first comfortable tier. There is no LiteRT shortcut for this model. |
| Why it is heavy | MatFormer nesting plus per-layer embeddings: the E4B name counts the effective configuration, but the file carries 7.8B parameters — the official-mirror weights total 14.6 GiB. The measured Q4_K_M (3.9 GiB) is what a 7.8B-total model should weigh, and the budget math follows the file, not the name. |
| Representative hardware | Galaxy A-class (6GB), iPhone 15/16 Pro tier (8GB); 12GB = Galaxy S Ultra / iPhone 17 Pro tier; 16GB = Pixel Pro tier; 24GB = OnePlus flagship tier |
| Source | bartowski GGUF listing + unsloth official mirror · accessed 2026-09-16 · measured |
⚠ Tight is a site-wide threshold: under 2.5 GiB of headroom. 8GB × Q4_K_M leaves 0.1 GiB (razor); 16GB × Q8_0 leaves 2.0 GiB (tight); 24GB is the comfort tier. On iOS the binding constraint is often the absolute per-app cap, not the percentage.
Tracked builds (actual file sizes, not theory)
| Build | Publisher | File size | 8GB phone | 12GB phone | Source |
|---|---|---|---|---|---|
| Q4_K_M (recommended) | bartowski | 3.95 GiB | TIGHT · 4.7 | TIGHT · 4.7 | HF @6dd3cb1e verified |
| Q4_K_M | unsloth | 4.23 GiB | NO · 5.0 | TIGHT · 5.0 | HF @90fa8b0e verified |
Q4_K_M is the sweet spot, and here the publisher IS the decision: bartowski's 3.95 GiB build undercuts unsloth's 4.23 GiB by 0.3 GiB on the same quant name — 7% apart, different tensor coverage, so they coexist rather than collapse. On an 8GB phone that gap is the difference between razor and no-fit (unsloth's totals 5.0 GiB against 4.8 GiB usable); on 12GB both fit tight. The official google repo is license-gated — files 403 with our token (metadata pinned), so this page tracks the community builds and the unsloth byte-mirror of the official weights (14.6 GiB, the BF16 anchor in the sources table).
How to run it on a phone
| iOS | PocketPal · LLM Farm |
| Android | PocketPal · llama.cpp (Termux or APK) |
| No LiteRT / No Gallery | litert-community does not ship a Gemma 3n build (negative existence, probed 2026-09-16), so Google AI Edge Gallery has no model file for it — GGUF loaders are the only tracked door. |
| Watch out | Thermal throttling cuts sustained tok/s on long generations; the OS reserve is why usable is 60%, not 100%; on iOS the per-app memory cap — not the percentage — is usually the binding constraint. |
Why phone RAM math is different
Phones have no swap: when memory runs out the app dies, not slows. In the community measurements we sampled, not one shows graceful degradation on a phone — iOS kills the app silently (jetsam), Android's LMK does the same; the boundary is a cliff, not a slope. KV cache shares one RAM pool with the OS and your background apps — that is why this page counts usable RAM at 60% of the sticker number. It is the same memory math as the GPU pages, with the budget switched from VRAM to unified phone RAM.
Running the same model on a desktop GPU is a different buying question — graphics-card VRAM, not the phone already in your pocket. For the other Google edge model already on a phone, see Gemma 4 E4B on a phone.
On thermals: short prompts feel instant almost everywhere the model fits. Sustained generation is where phones differ — a 2026 benchmark measured an iPhone 16 Pro dropping from ~40 tok/s peak to ~22.5 sustained over 20 iterations, and community reports put throttling onset at 30–60 seconds, so treat any speed number as a cold-start figure unless the source says otherwise (links in provenance).
The weights on this page are measured — GGUF file sizes read off the publishers' listings at pinned revisions (accessed 2026-09-16), with the official-caliber anchor from unsloth's byte-mirror of the gated google repo (14.6 GiB). What stays estimated is the total: KV cache at 4K context and a runtime reserve are added by simple arithmetic, and the KV figure is a classic-caliber upper bound derived from the pinned config, not a per-token measurement. The 0.60 usable rule stays an experience value — now with community measurements behind it — and if a better-supported figure shows up, the chart's zone boundaries move and the verdicts are recomputed.
Where every number comes from (every figure carries its tag)
Totals: labeled additive estimates — measured weights + KV allowance + runtime reserve.
| Tag | Slot | What it is | What the verdict can claim |
|---|---|---|---|
| measured | Weights | File sizes read off the publishers' Hugging Face listings at pinned revisions (accessed 2026-09-16): bartowski GGUF — Q2_K 2.57, IQ3_XS 2.95, Q4_K_M 3.95, Q5_K_M 4.61, Q6_K 5.31, Q8_0 6.85 GiB; unsloth GGUF — Q4_K_M 4.23 GiB (different build, 7% apart); unsloth official mirror — BF16-equivalent shards 14.62 GiB (google repo gated, metadata pinned c1221e9c6). Totals add a 4K KV allowance (0.27 GiB classic upper bound, from the pinned config: 2 KV heads × 256 dim × 35 layers) plus a 0.50 GiB runtime reserve. | Solid at file granularity — the weights are the files you will download; only the KV/reserve add-on is arithmetic. |
| reported | LiteRT absence | litert-community does not ship a Gemma 3n build: the author search returns 12 repos (Gemma 2/3/4 families) with no 3n entry (search result, probed 2026-09-16). Google AI Edge Gallery consequently has no model file for it. Listed as a negative-existence finding, not a build. | Stated plainly so you do not wait for a Gallery path that does not exist — re-checkable by the linked search. |
| reported | Usable = 60% rule | Experience value for the OS + background share, backed by community measurements: the maintained per-device table spans ~49–81% usable (per-device memory-budget table, accessed 2026-09-13), with a 2026 forums report of iPadOS 26.4 silently cutting 8GB iPads to ~37.5% (Apple Developer Forums 826325, accessed 2026-09-13 — single source, treated as provisional); iOS enforces a hard per-app cap rather than a percentage — PocketPal PR #651 measured 52% → 70% with the memory entitlement (accessed 2026-09-13). 0.60 is our conservative default inside that range. | Directionally safe — it separates 6GB phones from 24GB phones. On iOS the binding constraint is often the absolute per-app cap, not the percentage. |
| reported | App compatibility · thermal | PocketPal / LLM Farm / llama.cpp — project pages, not re-tested by us on every device; iPhone 16 Pro measured ~40 → ~22.5 tok/s over 20 iterations (2026 thermal benchmark paper, accessed 2026-09-13). | Expect app-loading to stand on the linked pages, and slowdown on long generations — never a cold-start speed promise. |
| bartowski/google_gemma-3n-E4B-it-GGUF | @6dd3cb1e | accessed 2026-10-12 |
| unsloth/gemma-3n-E4B-it-GGUF | @90fa8b0e | accessed 2026-10-12 |
| unsloth/gemma-3n-E4B-it (official byte-mirror) | 45e9fb1dd | accessed 2026-09-16 |
| google/gemma-3n-E4B-it (official, gated) | c1221e9c6 | metadata probed 2026-09-16 · files 403 |
| SO 5887248 · per-device memory budget table | latest | accessed 2026-09-13 |
| Apple Developer Forums 826325 · iPadOS 26.4 cap cut | latest | accessed 2026-09-13 |
| PocketPal PR #651 · entitlement 52%→70% | latest | accessed 2026-09-13 |
| 2026 thermal benchmark paper (arXiv 2603.23640) | 2026-03 | accessed 2026-09-13 |
FAQ
Can Gemma 3n E4B run on mobile?
Yes, via GGUF loaders: PocketPal or llama.cpp on Android, PocketPal or LLM Farm on iOS. There is no LiteRT build and no Google AI Edge Gallery path — litert-community does not ship one (negative existence, probed 2026-09-16).
Is 8GB RAM enough for Gemma 3n E4B?
Enough for the recommended Q4_K_M, razor-thin: 4.7 GiB total against 4.8 GiB usable, 0.1 GiB to spare. Take bartowski's 3.95 GiB build — unsloth's 4.23 GiB Q4_K_M totals 5.0 GiB and does not fit 8GB at all.
Is 6GB RAM enough for Gemma 3n E4B?
Only the Q2_K build: 3.3 GiB total against 3.6 GiB usable, tight. The recommended Q4_K_M wants 8GB. 4GB is out — the smallest tracked build totals 3.3 GiB.
Which Gemma 3n E4B build should I download?
bartowski's Q4_K_M (3.95 GiB) — it undercuts unsloth's same-name build (4.23 GiB) by 0.3 GiB, and on an 8GB phone that gap decides fit. At 6GB take Q2_K. There is no LiteRT build to choose.
Why is Gemma 3n E4B's file bigger than its name suggests?
MatFormer nesting plus per-layer embeddings: the E4B name counts the effective configuration, but the weights carry 7.8B parameters in total (the unsloth official mirror totals 14.6 GiB in BF16). The measured Q4_K_M reads 3.9 GiB — the file, not the name, sets the phone budget.