Gemma 3 27B
The previous Gemma generation, still very widely deployed and one of the best-documented models for local vision work. Its interleaved local/global attention keeps the KV cache smaller than the head count alone suggests.
Google DeepMindDensegeneral21 machines can run it
Total parameters
27.432B
Official spec sets memory need
Active parameters
27.432B
Official spec sets decode speed
Context
128K
Official spec tokens
4-bit weights
15 GB
Calculated before KV cache
Model specification
Publisher
Google DeepMind Official spec
Architecture
Dense transformer Official spec
Total parameters
27.432B Official spec
Context length
131,072 tokens Official spec
Attention ⓘ
GQA · 16 KV heads × 128 dim × 62 layers Official spec
Licence
Gemma Terms of Use Official spec source
Specification confidence ⓘ
High · verified 2026-09-07 Official spec source
Released
12 Mar 2025 Official spec
Official source
Quantizations and memory
| Quantization | Format | Bits/weight | Weights | Quality kept |
|---|---|---|---|---|
MLX 4-bit | mlx | 4.5 | 14.9 GB | 98.0% |
Q4_K_Mdefault | gguf | 4.85 | 16.0 GB | 98.5% |
Q8_0 | gguf | 8.5 | 27.7 GB | 99.9% |
BF16 | safetensors | 16 | 51.1 GB | 100.0% |
Weight sizes are computed from the parameter count and bits per weight plus a format-specific overhead for the layers that stay at higher precision — not read from a specific published file. Quality retention is an assumption, not a measured evaluation.
Recommendations
Cheapest that can run it
~1.3–1.8 t/s · €1.699
Lowest purchase price among configurations where the model fits at some quantization in our catalogue. Speed is not considered.
Cheapest above 20 t/s
~38.3–55.1 t/s · €2.299
20 tokens/second is roughly the point at which generation keeps pace with reading. Below it, interactive use feels like waiting.
Cheapest above 40 t/s
~38.3–55.1 t/s · €2.299
40 tokens/second is the threshold most people describe as comfortable for coding agents, where output arrives faster than you can review it.
Fastest with real measurements
Nothing in the database qualifies.
Highest throughput among configurations with an actual published measurement rather than our estimate.
Best throughput per euro
~65.5–94.2 t/s · €3.699
Highest decode tokens/second per EUR 1,000 of purchase price. Ignores running costs and resale — see the economics section for the full picture.
Hardware that runs Gemma 3 27B0 measured, 21 estimated
21 of 21 rows
| Hardware↕ | Quant | Memory↕ | Decode▼ | Prefill↕ | Context | Price↕ | Fit | Confidence↕ |
|---|---|---|---|---|---|---|---|---|
| Quad RTX 5090 workstation (4x 32 GB) NVIDIA · 128 GB · 7,168 GB/s | Q8_0 | 34.6 GB | ~76.4–110 t/s | ~3650–7590 t/s | 128K | €13.999 | Comfortable | Estimated |
| RTX 5090 workstation (1x 32 GB) NVIDIA · 32 GB · 1,792 GB/s | Q4_K_M | 21.3 GB | ~65.5–94.2 t/s | ~1120–2330 t/s | 16K | €3.699 | Comfortable | Estimated |
| Dual RTX 5090 workstation (2x 32 GB) NVIDIA · 64 GB · 3,584 GB/s | Q8_0 | 33.8 GB | ~51–73.4 t/s | ~1830–3790 t/s | 32K | €6.499 | Comfortable | Estimated |
| Dual used RTX 4090 workstation (48 GB) NVIDIA · 48 GB · 2,016 GB/s | Q4_K_M | 21.7 GB | ~49.7–71.5 t/s | ~2030–4220 t/s | 32K | €4.099 | Comfortable | Estimated |
| Used RTX 4090 workstation (24 GB) NVIDIA · 24 GB · 1,008 GB/s | Q4_K_M | 21.3 GB | ~38.3–55.1 t/s | ~1250–2600 t/s | 4K | €2.299 | Borderline | Estimated |
| RTX PRO 6000 Blackwell workstation (96 GB) NVIDIA · 96 GB · 1,792 GB/s | Q8_0 | 33.4 GB | ~36.1–52 t/s | ~1190–2460 t/s | 96K | €11.499 | Comfortable | Estimated |
| RTX PRO 6000 Max-Q workstation (96 GB, 300 W) NVIDIA · 96 GB · 1,792 GB/s | Q8_0 | 33.4 GB | ~36.1–52 t/s | ~971–2020 t/s | 96K | €11.999 | Comfortable | Estimated |
| Radeon AI PRO R9700 workstation (32 GB) AMD · 32 GB · 644 GB/s | Q4_K_M | 21.3 GB | ~29.3–42.2 t/s | ~896–1860 t/s | 16K | €2.799 | Comfortable | Estimated |
| Mac Studio M5 Ultra 512 GB Apple · 512 GB · 1,200 GB/s | Q8_0 | 33.4 GB | ~28.6–41.1 t/s | ~1060–2200 t/s | 128K | €12.599 | Comfortable | Estimated |
| Mac Studio M5 Ultra 256 GB Apple · 256 GB · 1,200 GB/s | Q8_0 | 33.4 GB | ~28.6–41.1 t/s | ~1060–2200 t/s | 128K | €9.599 | Comfortable | Estimated |
| Mac Studio M5 Ultra 96 GB Apple · 96 GB · 1,200 GB/s | Q8_0 | 33.4 GB | ~28.6–41.1 t/s | ~846–1760 t/s | 64K | €6.599 | Comfortable | Estimated |
| Mac Studio M5 Max 36 GB Apple · 36 GB · 614 GB/s | Q4_K_M | 21.3 GB | ~25.7–37 t/s | ~423–879 t/s | 16K | €2.999 | Comfortable | Estimated |
| Mac Studio M3 Ultra 512 GB Apple · 512 GB · 819 GB/s | Q8_0 | 33.4 GB | ~16–23 t/s | ~255–530 t/s | 128K | €9.199 | Comfortable | Estimated |
| Mac Studio M3 Ultra 256 GB Apple · 256 GB · 819 GB/s | Q8_0 | 33.4 GB | ~16–23 t/s | ~255–530 t/s | 128K | €6.299 | Comfortable | Estimated |
| Mac Studio M5 Max 128 GB Apple · 128 GB · 614 GB/s | Q8_0 | 33.4 GB | ~14.9–21.4 t/s | ~529–1100 t/s | 128K | €4.799 | Comfortable | Estimated |
| Mac Studio M5 Max 64 GB Apple · 64 GB · 614 GB/s | Q8_0 | 33.4 GB | ~14.9–21.4 t/s | ~529–1100 t/s | 32K | €3.899 | Comfortable | Estimated |
| Mac mini M4 Pro 64 GB Apple · 64 GB · 273 GB/s | Q8_0 | 33.4 GB | ~5.3–7.7 t/s | ~44.3–92 t/s | 32K | €2.499 | Comfortable | Estimated |
| NVIDIA DGX Spark 128 GB NVIDIA · 128 GB · 273 GB/s | Q8_0 | 33.4 GB | ~5.1–7.3 t/s | ~455–946 t/s | 128K | €4.299 | Comfortable | Estimated |
| Framework Desktop (Ryzen AI Max+ 395, 128 GB) AMD · 128 GB · 256 GB/s | Q8_0 | 33.4 GB | ~3.5–5 t/s | ~126–261 t/s | 128K | €2.399 | Comfortable | Estimated |
| GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB) AMD · 128 GB · 256 GB/s | Q8_0 | 33.4 GB | ~3–4.3 t/s | ~126–261 t/s | 128K | €1.899 | Comfortable | Estimated |
| CPU-only workstation (Ryzen 9950X, 192 GB DDR5) Generic · 192 GB · 90 GB/s | Q8_0 | 33.4 GB | ~1.3–1.8 t/s | ~15.2–31.5 t/s | 128K | €1.699 | Comfortable | Estimated |
Similar models
Gemma 4 31B
Google DeepMind · 31.273B · Gemma Terms of Use
Qwen3.8 27B
Alibaba Qwen · 27.781B · Apache 2.0
Gemma 4 26B-A4B
Google DeepMind · 25.806B (3.8B active) · Gemma Terms of Use
Mistral Small 3.2 24B
Mistral AI · 24.011B · Apache 2.0
Qwen3 32B
Alibaba Qwen · 32.762B · Apache 2.0
Qwen3 30B-A3B
Alibaba Qwen · 30.532B (3.3B active) · Apache 2.0