Gemma 3 27B
Google's dense 27B multimodal model — text and image input, a 128K context window, and a 16.5 GB Q4 footprint that fits comfortably on any 24 GB card.
Last verified 2026-07 · Page created; Q4_K_M/Q5_K_M/Q8_0 sizes verified against bartowski/google_gemma-3-27b-it-GGUF; full-precision size is the published BF16 file.
What it is good at
Gemma 3 27B is Google's largest openly-released Gemma 3 model: dense, multimodal (it accepts image input alongside text), and trained across 140+ languages with a 128K context window — the combination of long context and vision input in a single 27B checkpoint is unusual for this size class.
It ships under Google's Gemma license rather than Apache or MIT — permissive for self-hosting and commercial use, but read the actual terms if you are shipping a product on top of it rather than assuming it behaves identically to Apache 2.0.
The hardware it wants
Q4_K_M is 16.55 GB, which fits a 24 GB GPU with plenty of room to spare, and is workable even on a 20 GB card. Q5_K_M (19.27 GB) still fits 24 GB comfortably. Q8_0 (28.71 GB) needs a 32 GB+ card or two pooled 24 GB GPUs. The full-precision weights are published as BF16 at 54.03 GB (same 2-byte-per-parameter size as FP16), useful mainly as a fine-tuning starting point rather than for inference.
VRAM by quantization
| Quant | File size | VRAM @ 8K ctx | VRAM @ 32K ctx | 8 GB | 16 GB | 24 GB | 32 GB | 48 GB |
|---|---|---|---|---|---|---|---|---|
| FP16 | 54.0 GB | 66.0 GB | 77.7 GB | Doesn't fit | Doesn't fit | Doesn't fit | Doesn't fit | Doesn't fit |
| Q8_0 | 28.7 GB | 36.9 GB | 48.5 GB | Doesn't fit | Doesn't fit | Doesn't fit | Doesn't fit | Fits comfortably |
| Q5_K_M | 19.3 GB | 26.1 GB | 37.7 GB | Doesn't fit | Doesn't fit | Tight — reduced context | Fits comfortably | Fits comfortably |
| Q4_K_M | 16.6 GB | 22.9 GB | 34.6 GB | Doesn't fit | Doesn't fit | Fits comfortably | Fits comfortably | Fits comfortably |
Fit verdicts use the 8K context preset. Estimated from published weight sizes plus standard KV-cache math — see methodology.
Runs on
Unified-memory machines — Apple Silicon Mac Studio/MacBook Pro (Max/Ultra chips) and AMD Strix Halo mini-PCs — share system RAM and GPU memory in one pool, so a 64–128 GB configuration can hold models that would otherwise need multiple discrete GPUs. Throughput is bandwidth-bound like any GPU; it isn’t modeled in the table above yet.
Recommended builds
Three ways to size a machine for Gemma 3 27B, from the smallest GPU that fits to the no-compromise option with headroom for the next model up.
Tokens/sec expectations
| GPU | Bandwidth | Est. tok/s (Q4_K_M) |
|---|---|---|
| RTX 5090 | 1792 GB/s | ~75.8 |
| RTX 4090 | 1008 GB/s | ~42.6 |
| RTX 6000 Ada | 960 GB/s | ~40.6 |
| RX 7900 XTX | 960 GB/s | ~40.6 |
| RTX 5080 | 960 GB/s | ~40.6 |
| RTX 5070 Ti | 896 GB/s | ~37.9 |
Estimated from memory bandwidth ÷ active weight size — a ceiling, not a measured figure. Real throughput depends on the inference engine, batch size, and context length.
Build it, or rent the compute?
Owning the hardware is a one-time cost that keeps paying off the more you actually use it — every run after the first is free, the data never leaves your machine, and the GPU still does everything else a PC does (gaming, editing, the rest of your workload) between AI sessions. Renting cloud GPU time trades that upfront cost for pay-as-you-go pricing with no idle hardware to maintain, which wins for occasional, bursty, or much-larger-than-consumer-VRAM jobs. For a model that fits comfortably in the VRAM tiers above and gets used regularly, owning the GPU tends to be the cheaper path over time — for anything larger than a single card can hold, renting is often the more practical starting point.
Frequently asked questions
What GPU do I need for Gemma 3 27B?
A single 24 GB card (RTX 3090/4090) handles Q4_K_M or Q5_K_M comfortably. A 16 GB card can run it at a more aggressive quant with some quality loss.
Does Gemma 3 27B actually understand images, or is that a separate model?
It is built in — the same checkpoint accepts image input alongside text, so you do not need a separate vision adapter or model for basic image understanding tasks.
Is the Gemma license the same as Apache 2.0?
No — it is Google's own Gemma license. It permits commercial use and self-hosting, but it is a distinct set of terms, not Apache 2.0; check the license text if you are building a commercial product on top of it.