| Model profile | ||
|---|---|---|
| Model | Size | Approximate size |
| Gemma 4 12BModel source: provider262,144 tokens max native context | 12B | 7.2 GB4-bit quantization |
| Compatible hardware Estimated speeds, not benchmark results: calculated from memory bandwidth and model size. Real results can differ significantly because there is no precise formula for deriving LLM generation speed from hardware specifications alone. | ||||
|---|---|---|---|---|
| Hardware | Memory | Bandwidth | vLLM | Estimated generation |
| Raspberry Pi 5 | 16 GBLPDDR4X RAM | 17 GB/s | no | 1 tok/s4-bit quantization |
| DDR4 16GB Computer | 16 GBRAM | 51.2 GB/s | no | 3 tok/s4-bit quantization |
| DDR5 16GB Computer | 16 GBRAM | 89.6 GB/s | no | 6 tok/s4-bit quantization |
| DDR5 32GB Computer | 32 GBRAM | 89.6 GB/s | no | 6 tok/s4-bit quantization |
| Mac mini M4 | 24 GBUnified RAM | 120 GB/s | no | 9 tok/s4-bit quantization |
| Mac mini M6 | 24 GBUnified RAM | 170 GB/s | no | 12 tok/s4-bit quantization |
| RTX 2000 Ada | 16 GBGPU VRAM | 224 GB/s | yes | 16 tok/s4-bit quantization |
| Radeon 8060S 96GB | 96 GBUnified RAM | 256 GB/s | yes | 19 tok/s4-bit quantization |
| M4 Pro | 24 GBUnified RAM | 273 GB/s | no | 20 tok/s4-bit quantization |
| DGX Spark | 128 GBUnified RAM | 273 GB/s | yes | 20 tok/s4-bit quantization |
| RTX 4060 Ti 16GB | 16 GBGPU VRAM | 288 GB/s | yes | 21 tok/s4-bit quantization |
| M5 Pro | 64 GBUnified RAM | 307 GB/s | no | 23 tok/s4-bit quantization |
| RTX 3060 12GB | 12 GBGPU VRAM | 360 GB/s | yes | 26 tok/s4-bit quantization |
| RTX 5060 Ti 16GB | 16 GBGPU VRAM | 448 GB/s | yes | 33 tok/s4-bit quantization |
| M5 Max | 128 GBUnified RAM | 614 GB/s | no | 45 tok/s4-bit quantization |
| M3 Ultra | 96 GBUnified RAM | 819 GB/s | no | 59 tok/s4-bit quantization |
| RTX 5080 | 16 GBGPU VRAM | 960 GB/s | yes | 69 tok/s4-bit quantization |
| RTX 3090 Ti | 24 GBGPU VRAM | 1008 GB/s | yes | 72 tok/s4-bit quantization |
| M5 Ultra | 256 GBUnified RAM | 1200 GB/s | no | 85 tok/s4-bit quantization |
| RTX 5090 | 32 GBGPU VRAM | 1792 GB/s | yes | 124 tok/s4-bit quantization |
| RTX PRO 6000 | 96 GBGPU VRAM | 1792 GB/s | yes | 124 tok/s4-bit quantization |
| DGX H200 | 1128 GBGPU HBM3e | 4800 GB/s | yes | 287 tok/s4-bit quantization |
| DGX Station | 748 GBCoherent Memory | 7100 GB/s | yes | 385 tok/s4-bit quantization |
| ET900N G3 | 748 GBCoherent Memory | 7100 GB/s | yes | 385 tok/s4-bit quantization |
| DGX B200 | 1440 GBGPU HBM3e | 8000 GB/s | yes | 419 tok/s4-bit quantization |
| GB200 NVL72 | 13400 GBGPU HBM3e | 8000 GB/s | yes | 419 tok/s4-bit quantization |
| GB300 NVL72 | 20000 GBGPU HBM3e | 8000 GB/s | yes | 419 tok/s4-bit quantization |
Demonstrates running Claude Code locally on Windows via Ollama using models like Qwen 3.8 27B, covering installation, configuration, and usage examples such as writing expenses.py.
Installs Ollama on Windows, runs models locally via PowerShell, tests CPU performance, manages libraries, adjusts settings, and demonstrates offline image analysis.
Token generation speeds were measured for Gemma 4 E2B, Gemma 4 12B, gpt-oss 20B, Gemma 4 26B, Qwen 3.6 35B, and Qwen 3.8 27B using Ollama on a Geekom A9 Max mini PC with 32 GB RAM.
Benchmark IBM Granite 4.2, Qwen 3.8 27B, Gemma 4, and Ornith 1.5 on an RTX 5090, measuring KV cache cost, usable context, generation speed, coding, math, and tool calling via Python tests.
Tests Gemma 4 12B Coder locally on MacBook M4 Pro using Ollama and Pi agent via color palette generator, pixel art editor single prompts, and step-by-step development across 18 tasks.
Qwen, Ornith, Qwythos, and Gemma 4 compete in a two-round battle using easy and challenging prompts to demonstrate the importance of writing a good prompt.
Tests Bonsai 27B binary and ternary variants against Qwen 3.6 35B-A3B and Gemma 4 12B on an RTX 3060 using llama.cpp, measuring generation speed, total memory footprint, and performance on web design
Tests major Gemma 4 update with QAT GGUF, covering tool-calling fix, FA4 support, and vision token control.
Tests Ollama and Llama.cpp running inside Docker on ZimaBoard 2, evaluating speed using Gemma-4 with Vision and GPT-OSS from OpenAI, plus an endurance thermal throttle check.
Qwen 3.6 and Gemma 4 variants are tested to design and build a full-screen visualizer using MusicPlayer, with results hosted in the ext-mp-ai-visualizers repo.
Tests Ornith-1.0-9B versus Gemma-4-12B in local AI coding tasks to evaluate whether Ornith's benchmark scores align with real-world performance.
Tests Gemma 4 12B finetune with Fable 5 and Composer 2.5 against base via Q4_K_M quant on 16GB VRAM using Performance, Memory, Agency, OpenAI Human Eval, Sand Physics and Dungeon Crawler benchmarks.
Compares Google QAT and Unsloth QAT + MTP using identical QAT base weights and MTP setup to evaluate performance differences where the quantization method is the sole variable.
Compares Gemma4 12B and Gemma4 12B QAT via VPS Dashboard Setup, Tower Defense Game addition, and Chat Client Build to evaluate instruction following, UI cleanliness, reliability, and usability under f
A bug fix in Gemma 4's chat template resolves issues that were silently degrading multi-turn agent performance.
Compares Gemma4 12B and Gemma4 12B QAT via iOS UI Clone, Weather Dashboard, and Broken Code Debug + Repair Challenge using single-file HTML, CSS, and JavaScript tests.
Demonstrates running Gemma 4 12B QAT locally via LM Studio on Windows PCs with 8GB VRAM, integrated into VS Code using the Continue extension for offline coding assistance.
Compares Unsloth’s q4 QAT and regular q4 GGUF variants of Gemma 4 12B on adherence, agency, coding, and memory using a 16GB VRAM system.
Stacks Google's QAT quantization with llama.cpp's MTP support to run Gemma 4 12B at double the speed locally.
Tests Google QAT versus Unsloth Q4_0 quantized versions of Gemma 4 12B at identical file sizes to determine which approach yields better results.
Tests Gemma 4 12B q8 vs q4 quantizations on a system with 16GB VRAM and 32GB DDR4 RAM across performance, memory, coding, and agency benchmarks using unsloth/gemma-4-12b-it-GGUF.
Tests Gemma 4 12B's encoder-free multimodal architecture, coding, reasoning, and Three.js generation via World of AI Bench, comparing it against Qwen3.6-35B-A3B.
Locally installs and tests Gemma 4 12B QAT via Ollama using coding demos, multilingual translation, creativity, and situation demonstrations.
Evaluates instruction adherence, UI cleanliness, functionality, and interactivity using single-file HTML/CSS/JavaScript tests: iPhone Replica, Top-Down Car Game, and Live Weather Dashboard.
Evaluates unsloth/gemma-4-12b-it-GGUF on a system with 16GB VRAM and 32GB DDR4 RAM across memory, agency, reasoning, implementation, audio and vision tasks.
Compares Gemma4 12B against Qwen3.6 27B using the same GPU setup.
Tests Gemma-4-12b-it vision, security via Splunk CTF and AWS S3 investigation with Anvor AI, tool-calling, and building a game with Pi on NVIDIA or AMD GPUs.
Tests Gemma 4 12B via local setup config and practical evaluations including browser OS, 3D printer simulation, image to SVG, scene generation, multimodal websites, wireframe to site, and OpenCode tas
Locally installed Gemma 4 12B encoder-free model demonstrates unified architecture through text inference, coding, multilingual translation, image understanding, OCR, and audio transcription demos.