Gemma 4 12B hardware requirements
What is Gemma 4 12B?
Gemma 4 12B is a dense model that Google DeepMind added to the Gemma 4 family in June 2026. Google calls it unified because it has no separate vision or audio encoder; it takes text, images and short audio clips with up to a 256k context. It targets reasoning, coding and function calling, its thinking mode is off by default, and the license is Apache 2.0.
About the model
- Parameters
- 11.96B
- Attention
- Sliding-window attention
- Max context
- 262,144 tokens
- Released
- 2026-05-23
- Model card
- Gemma 4 12B on Hugging Face
Other sizes in this series
| Model | Parameters | Baseline quant | Memory needed (8k) | GeForce RTX 4060 | GeForce RTX 5060 Ti 16GB | GeForce RTX 3090 |
|---|---|---|---|---|---|---|
| Gemma 4 E2B | 5.12B | Q4_K_M | 2.1 GB | Runs great | Runs great | Runs great |
| Gemma 4 E4B | 8B | Q4_K_M | 3.7 GB | Runs great | Runs great | Runs great |
| Gemma 4 12Bthis model | 11.96B | Q4_K_M | 7.8 GB | Runs well | Runs great | Runs great |
| Gemma 4 26B-A4B | 26.54B total · 4B active (MoE) | Q4_K_M | 17.7 GB | Runs well | Runs well | Runs great |
| Gemma 4 31B | 31.27B | Q4_K_M | 19.5 GB | Runs slowly | Runs slowly | Runs great |
GGUF files
| Quant | File size | Quality | Published by |
|---|---|---|---|
| UD-Q2_K_XL | 4.66 GB | Heavy loss | Community GGUF · unsloth |
| IQ4_XS | 6.38 GB | Good balance | Community GGUF · unsloth |
| Q4_K_Mbaseline | 7.12 GB | Good balance | Community GGUF · unsloth |
| Q8_0 | 12.67 GB | Near-lossless | Community GGUF · unsloth |
File sizes do not include the vision projector (mmproj) that image input needs.
Some repositories publish only Unsloth Dynamic (UD) variants; sizes are those of the file you would download.
Memory needed by quant and context
| Quant | 4k | 8k | 16k | 32k | 64k | 128k | 256k |
|---|---|---|---|---|---|---|---|
| UD-Q2_K_XL | 5.2 | 5.3 | 5.6 | 6.0 | 6.8 | 8.5 | 11.8 |
| IQ4_XS | 7.0 | 7.1 | 7.3 | 7.7 | 8.5 | 10.2 | 13.6 |
| Q4_K_M | 7.7 | 7.8 | 8.0 | 8.4 | 9.3 | 10.9 | 14.3 |
| Q8_0 | 13.3 | 13.4 | 13.6 | 14.0 | 14.8 | 16.5 | 19.8 |
Which GPUs can run Gemma 4 12B?
How to read the verdicts
- Runs great
- Fully on the GPU at 20 tok/s or more
- Runs well
- Fully on the GPU at 8–20 tok/s; MoE experts in system RAM at 20 tok/s or more; or at least 90% on the GPU at 8 tok/s or more
- Runs slowly
- 2–8 tok/s; CPU-only; less than 90% on the GPU; or MoE experts in system RAM below 20 tok/s
- Won't run
- Does not fit, or under 2 tok/s
Dense models split between GPU and CPU slow down sharply — the CPU side sets the pace. MoE models that keep only their experts in system RAM degrade far more gently.
Filter this list by budget, type and context in the calculator
Scroll sideways to see every column.
| Hardware | Verdict | Speed | Quant | Memory | Runs as | Context | Notes | Price |
|---|---|---|---|---|---|---|---|---|
| GeForce RTX 5090 | est. 172.9 tok/s152.2–193.7calibrated estimate ±12% | Q4_K_M | 8.4 / 32.0 GB | Full GPU | up to 256k | — | $4,200 street (as of 2026-08-09) | |
| NVIDIA RTX PRO 6000 Blackwell | Runs great | est. 172.9 tok/s152.2–193.7calibrated estimate ±12% | Q4_K_M | 8.4 / 96.0 GB | Full GPU | up to 256k | — | — |
| NVIDIA RTX PRO 6000 Blackwell Max-Q | Runs great | est. 172.9 tok/s152.2–193.7calibrated estimate ±12% | Q4_K_M | 8.4 / 96.0 GB | Full GPU | up to 256k | — | — |
| Radeon RX 7900 XTX | Runs great | est. 100.6 tok/s88.5–112.6calibrated estimate ±12% | Q4_K_M | 8.4 / 24.0 GB | Full GPU | up to 256k | — | $999 launch MSRP |
| GeForce RTX 3090 Ti | Runs great | est. 97.3 tok/s85.6–108.9calibrated estimate ±12% | Q4_K_M | 8.4 / 24.0 GB | Full GPU | up to 256k | — | $1,999 launch MSRP |
| GeForce RTX 4090 | est. 97.3 tok/s85.6–108.9calibrated estimate ±12% | Q4_K_M | 8.4 / 24.0 GB | Full GPU | up to 256k | — | $2,755 street (as of 2026-08-09) | |
| Apple M5 Ultra · 96 GB | Runs great | est. 96.5 tok/s77.2–115.8calibrated estimate ±20% | Q4_K_M | 7.8 / 67.2 GB | Unified memory | up to 256k | — | — |
| GeForce RTX 5080 | Runs great | est. 92.6 tok/s81.5–103.8calibrated estimate ±12% | Q4_K_M | 8.4 / 16.0 GB | Full GPU | up to 256k | — | $1,256 street (as of 2026-08-09) |
| GeForce RTX 3090 | est. 90.3 tok/s79.5–101.2calibrated estimate ±12% | Q4_K_M | 8.4 / 24.0 GB | Full GPU | up to 256k | — | $1,050 street (as of 2026-08-09) | |
| GeForce RTX 3080 12GB | Runs great | est. 88.0 tok/s77.4–98.6calibrated estimate ±12% | Q4_K_M | 8.4 / 12.0 GB | Full GPU | up to 128k | — | $799 launch MSRP |
| GeForce RTX 5070 Ti | Runs great | est. 86.5 tok/s76.1–96.8calibrated estimate ±12% | Q4_K_M | 8.4 / 16.0 GB | Full GPU | up to 256k | — | $949 street (as of 2026-08-09) |
| GeForce RTX 5080 Laptop | Runs great | est. 86.5 tok/s76.1–96.8calibrated estimate ±12% | Q4_K_M | 8.4 / 16.0 GB | Full GPU | up to 256k | — | — |
| GeForce RTX 5090 Laptop | Runs great | est. 86.5 tok/s76.1–96.8calibrated estimate ±12% | Q4_K_M | 8.4 / 24.0 GB | Full GPU | up to 256k | — | — |
| Radeon RX 7900 XT | Runs great | est. 83.8 tok/s73.8–93.9calibrated estimate ±12% | Q4_K_M | 8.4 / 20.0 GB | Full GPU | up to 256k | — | $899 launch MSRP |
| GeForce RTX 3080 10GB | Runs great | est. 73.3 tok/s64.5–82.1calibrated estimate ±12% | Q4_K_M | 8.4 / 10.0 GB | Full GPU | up to 64k | — | $699 launch MSRP |
| GeForce RTX 4080 Super | Runs great | est. 71.0 tok/s62.5–79.5calibrated estimate ±12% | Q4_K_M | 8.4 / 16.0 GB | Full GPU | up to 256k | — | $999 launch MSRP |
| GeForce RTX 4080 | Runs great | est. 69.2 tok/s60.9–77.5calibrated estimate ±12% | Q4_K_M | 8.4 / 16.0 GB | Full GPU | up to 256k | — | $1,199 launch MSRP |
| Radeon RX 9070 | Runs great | est. 67.6 tok/s59.5–75.7calibrated estimate ±12% | Q4_K_M | 8.4 / 16.0 GB | Full GPU | up to 256k | — | $639 street (as of 2026-08-09) |
| Radeon RX 9070 XT | Runs great | est. 67.6 tok/s59.5–75.7calibrated estimate ±12% | Q4_K_M | 8.4 / 16.0 GB | Full GPU | up to 256k | — | $689 street (as of 2026-08-09) |
| Radeon RX 7800 XT | Runs great | est. 65.4 tok/s57.5–73.2calibrated estimate ±12% | Q4_K_M | 8.4 / 16.0 GB | Full GPU | up to 256k | — | $499 launch MSRP |
| GeForce RTX 4070 Ti Super | Runs great | est. 64.8 tok/s57.1–72.6calibrated estimate ±12% | Q4_K_M | 8.4 / 16.0 GB | Full GPU | up to 256k | — | $799 launch MSRP |
| GeForce RTX 5070 | est. 64.8 tok/s57.1–72.6calibrated estimate ±12% | Q4_K_M | 8.4 / 12.0 GB | Full GPU | up to 128k | — | $629 street (as of 2026-08-09) | |
| GeForce RTX 5070 Ti Laptop | Runs great | est. 64.8 tok/s57.1–72.6calibrated estimate ±12% | Q4_K_M | 8.4 / 12.0 GB | Full GPU | up to 128k | — | — |
| Apple M3 Ultra · 96 GB | Runs great | est. 64.4 tok/s51.5–77.2calibrated estimate ±20% | Q4_K_M | 7.8 / 67.2 GB | Unified memory | up to 256k | — | — |
| Apple M5 Max (40-core GPU) · 48 GB | Runs great | est. 64.3 tok/s51.5–77.2calibrated estimate ±20% | Q4_K_M | 7.8 / 33.6 GB | Unified memory | up to 256k | — | — |
| Apple M2 Ultra · 64 GB | Runs great | est. 62.9 tok/s50.3–75.4calibrated estimate ±20% | Q4_K_M | 7.8 / 44.8 GB | Unified memory | up to 256k | — | — |
| GeForce RTX 2080 Ti | Runs great | est. 59.4 tok/s52.3–66.6calibrated estimate ±12% | Q4_K_M | 8.4 / 11.0 GB | Full GPU | up to 64k | — | $999 launch MSRP |
| Apple M4 Max (40-core GPU) · 48 GB | Runs great | est. 57.2 tok/s45.8–68.6calibrated estimate ±20% | Q4_K_M | 7.8 / 33.6 GB | Unified memory | up to 256k | — | — |
| GeForce RTX 4090 Laptop | Runs great | est. 55.6 tok/s48.9–62.3calibrated estimate ±12% | Q4_K_M | 8.4 / 16.0 GB | Full GPU | up to 256k | — | — |
| GeForce RTX 4070 | est. 48.6 tok/s42.8–54.5calibrated estimate ±12% | Q4_K_M | 8.4 / 12.0 GB | Full GPU | up to 128k | — | $599 launch MSRP | |
| GeForce RTX 4070 Super | Runs great | est. 48.6 tok/s42.8–54.5calibrated estimate ±12% | Q4_K_M | 8.4 / 12.0 GB | Full GPU | up to 128k | — | $599 launch MSRP |
| Apple M5 Max (32-core GPU) · 36 GB | Runs great | est. 48.2 tok/s38.6–57.8calibrated estimate ±20% | Q4_K_M | 7.8 / 25.2 GB | Unified memory | up to 256k | — | — |
| GeForce GTX 1080 Ti | Runs great | est. 46.7 tok/s32.7–60.7theoretical estimate ±30% | Q4_K_M | 8.4 / 11.0 GB | Full GPU | up to 64k | — | $699 launch MSRP |
| GeForce RTX 5060 Ti 16GB | est. 43.2 tok/s38.0–48.4calibrated estimate ±12% | Q4_K_M | 8.4 / 16.0 GB | Full GPU | up to 128k | — | $569 street (as of 2026-08-09) | |
| Apple M4 Max (32-core GPU) · 36 GB | Runs great | est. 43.0 tok/s34.4–51.5calibrated estimate ±20% | Q4_K_M | 7.8 / 25.2 GB | Unified memory | up to 128k | — | — |
| GeForce RTX 4080 Laptop | Runs great | est. 41.7 tok/s36.7–46.7calibrated estimate ±12% | Q4_K_M | 8.4 / 12.0 GB | Full GPU | up to 128k | — | — |
| Intel Arc B580 | Runs great | 41.3 tok/s8k estimate 37.1 tok/s (29.7–44.5, calibrated estimate ±20%)measured (1 run, 512 context) | Q4_K_M | 8.4 / 12.0 GB | Full GPU | up to 64k | — | $290 street (as of 2026-08-09) |
| Radeon RX 9060 XT 16GB | Runs great | 38.3 tok/s8k estimate 33.7 tok/s (29.7–37.8, calibrated estimate ±12%)measured (2 runs, 512 context) | Q4_K_M | 8.4 / 16.0 GB | Full GPU | up to 32k | — | $449 street (as of 2026-08-09) |
| GeForce RTX 3060 12GB | est. 34.7 tok/s30.6–38.9calibrated estimate ±12% | Q4_K_M | 8.4 / 12.0 GB | Full GPU | up to 64k | — | $250 street (as of 2026-08-09) | |
| Apple M5 Pro · 24 GB | Runs great | est. 32.2 tok/s25.7–38.6calibrated estimate ±20% | Q4_K_M | 7.8 / 16.8 GB | Unified memory | up to 32k | — | — |
| GeForce RTX 3070 | Runs well | est. 31.7 tok/s25.4–38.1calibrated estimate ±20% | Q4_K_M | 8.0 / 8.0 GB + 0.4 GB RAM | Partial offload | up to 16k |
| $499 launch MSRP |
| GeForce RTX 5060 | Runs well | est. 31.7 tok/s25.4–38.1calibrated estimate ±20% | Q4_K_M | 8.0 / 8.0 GB + 0.4 GB RAM | Partial offload | up to 16k |
| $339 street (as of 2026-08-09) |
| GeForce RTX 5060 Ti 8GB | Runs well | est. 31.7 tok/s25.4–38.1calibrated estimate ±20% | Q4_K_M | 8.0 / 8.0 GB + 0.4 GB RAM | Partial offload | up to 16k |
| $429 street (as of 2026-08-09) |
| NVIDIA DGX Spark · 128 GB | Runs well | est. 29.4 tok/s23.5–35.2calibrated estimate ±20% | Q4_K_M | 7.8 / 116.9 GB | Unified memory | up to 256k |
| $3,999 launch MSRP (128 GB) |
| Apple M4 Pro · 24 GB | est. 28.6 tok/s22.9–34.3calibrated estimate ±20% | Q4_K_M | 7.8 / 16.8 GB | Unified memory | up to 256k |
| — | |
| GeForce RTX 4060 Ti 16GB | Runs well | est. 27.8 tok/s24.5–31.1calibrated estimate ±12% | Q4_K_M | 8.4 / 16.0 GB | Full GPU | up to 256k |
| $499 launch MSRP |
| Ryzen AI Max+ 395 (Strix Halo) · 32 GB | Runs well | 24.4 tok/s8k estimate 28.9 tok/s (23.2–34.7, calibrated estimate ±20%)measured (1 run, 512 context) | Q4_K_M | 7.8 / 22.4 GB | Unified memory | up to 256k |
| $1,999 launch MSRP (128 GB) |
| GeForce RTX 4060 Ti 8GB | Runs well | est. 22.9 tok/s18.4–27.5calibrated estimate ±20% | Q4_K_M | 8.0 / 8.0 GB + 0.4 GB RAM | Partial offload | up to 16k |
| $399 launch MSRP |
| GeForce RTX 4060 | est. 21.9 tok/s17.6–26.3calibrated estimate ±20% | Q4_K_M | 8.0 / 8.0 GB + 0.4 GB RAM | Partial offload | up to 16k |
| $299 launch MSRP | |
| GeForce RTX 4060 Laptop | Runs well | est. 20.9 tok/s16.7–25.1calibrated estimate ±20% | Q4_K_M | 8.0 / 8.0 GB + 0.4 GB RAM | Partial offload | up to 16k |
| — |
| GeForce RTX 4070 Laptop | Runs well | est. 20.9 tok/s16.7–25.1calibrated estimate ±20% | Q4_K_M | 8.0 / 8.0 GB + 0.4 GB RAM | Partial offload | up to 16k |
| — |
| GeForce RTX 3050 8GB | Runs well | est. 18.8 tok/s15.0–22.5calibrated estimate ±20% | Q4_K_M | 8.0 / 8.0 GB + 0.4 GB RAM | Partial offload | up to 16k |
| $249 launch MSRP |
| Apple M5 · 16 GB | Runs well | est. 16.1 tok/s12.9–19.3calibrated estimate ±20% | Q4_K_M | 7.8 / 11.2 GB | Unified memory | up to 128k | — | — |
| Apple M6 · 16 GB | Runs well | est. 16.1 tok/s12.9–19.3calibrated estimate ±20% | Q4_K_M | 7.8 / 11.2 GB | Unified memory | up to 128k | — | — |
| Apple M4 · 16 GB | Runs wellGemma 4 12B verdict on the Apple M4 | est. 12.6 tok/s10.1–15.1calibrated estimate ±20% | Q4_K_M | 7.8 / 11.2 GB | Unified memory | up to 16k | — | — |
| DDR5-6000 dual-channel CPU · 16 GB | Runs slowly | est. 6.6 tok/s5.3–7.9calibrated estimate ±20% | Q4_K_M | 7.3 GB RAM | CPU only | up to 256k | — | — |
| DDR5-5600 dual-channel CPU · 16 GB | Runs slowly | est. 6.2 tok/s4.9–7.4calibrated estimate ±20% | Q4_K_M | 7.3 GB RAM | CPU only | up to 256k | — | — |
| DDR4-3200 dual-channel CPU · 16 GB | Runs slowly | est. 3.5 tok/s2.8–4.2calibrated estimate ±20% | Q4_K_M | 7.3 GB RAM | CPU only | up to 64k | — | — |
Models with a thinking mode spend extra tokens thinking, so their speed thresholds are 1.5× stricter (30 / 12 / 3 tok/s). With thinking turned off, a model may do one grade better.
Cheapest GPUs that run it
- Runs greatGeForce RTX 3060 12GB$250 street (as of 2026-08-09), est. 34.7 tok/s (30.6–38.9, calibrated estimate ±12%).
Measured results for Gemma 4 12B
Public benchmarks we calibrate against. Their conditions (context, backend, flags) can differ from the estimates above.
| Hardware | Quant | Backend | Context | Prompt (tok/s) | Generation (tok/s) | Flags | Source | Measured |
|---|---|---|---|---|---|---|---|---|
| Radeon RX 9060 XT 16GB | Q4_K_M | llama.cpp | 512 | 1,509.0 | 39.0 | Vulkan | github.com | 2026-10-02 |
| Radeon RX 9060 XT 16GB | Q4_K_M | llama.cpp | 512 | 1,467.0 | 37.6 | ROCm | github.com | 2026-10-02 |
| Intel Arc B580 | Q4_K_M | llama.cpp | 512 | 1,320.0 | 41.3 | SYCL, FA off | github.com | 2026-06-13 |
| Ryzen AI Max+ 395 (Strix Halo) | IQ4_XS | llama.cpp | 512 | 680.0 | 25.7 | Vulkan | github.com | 2026-06-04 |
| Ryzen AI Max+ 395 (Strix Halo) | Q4_K_M | llama.cpp | 512 | 685.0 | 24.4 | Vulkan | github.com | 2026-06-04 |
Frequently asked questions
How much VRAM does Gemma 4 12B need?
At Q4_K_M the file is 7.12 GB; with 8k context a graphics card needs about 7.8 GB (weights 7.10 GB, KV cache 0.1 GB, compute buffer 0.6 GB). The smallest tracked file (UD-Q2_K_XL (heavy quality loss)) needs about 5.3 GB.
What is the cheapest GPU that runs Gemma 4 12B well?
Cheapest that runs it great: GeForce RTX 3060 12GB at $250 (street price as of 2026-08-09), est. 34.7 tok/s (30.6–38.9, calibrated estimate ±12%).
Can Gemma 4 12B run on an 8 GB, 16 GB or 24 GB GPU?
At Q4_K_M with 8k context: GeForce RTX 4060 — Runs well, est. 21.9 tok/s (17.6–26.3, calibrated estimate ±20%); GeForce RTX 5060 Ti 16GB — Runs great, est. 43.2 tok/s (38.0–48.4, calibrated estimate ±12%); GeForce RTX 3090 — Runs great, est. 90.3 tok/s (79.5–101.2, calibrated estimate ±12%).
Speeds are estimates from memory bandwidth, calibrated against public benchmarks, and each one comes with an error band and a confidence label. Real results vary with drivers, backend, context length and thermals.