| Model profile | ||
|---|---|---|
| Model | Size | Approximate size |
| Gemma 4 26BModel source: provider262,144 tokens max native context | 26Bactive 4B | 15.6 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 |
| DDR5 32GB Computer | 32 GBRAM | 89.6 GB/s | no | 15 tok/s4-bit quantization |
| Mac mini M4 | 24 GBUnified RAM | 120 GB/s | no | 20 tok/s4-bit quantization |
| Radeon 8060S 96GB | 96 GBUnified RAM | 256 GB/s | yes | 41 tok/s4-bit quantization |
| M4 Pro | 24 GBUnified RAM | 273 GB/s | no | 44 tok/s4-bit quantization |
| DGX Spark | 128 GBUnified RAM | 273 GB/s | yes | 44 tok/s4-bit quantization |
| M5 Pro | 64 GBUnified RAM | 307 GB/s | no | 49 tok/s4-bit quantization |
| M5 Max | 128 GBUnified RAM | 614 GB/s | no | 92 tok/s4-bit quantization |
| M3 Ultra | 96 GBUnified RAM | 819 GB/s | no | 117 tok/s4-bit quantization |
| RTX 3090 Ti | 24 GBGPU VRAM | 1008 GB/s | yes | 139 tok/s4-bit quantization |
| M5 Ultra | 256 GBUnified RAM | 1200 GB/s | no | 159 tok/s4-bit quantization |
| RTX 5090 | 32 GBGPU VRAM | 1792 GB/s | yes | 213 tok/s4-bit quantization |
| RTX PRO 6000 | 96 GBGPU VRAM | 1792 GB/s | yes | 213 tok/s4-bit quantization |
| DGX H200 | 1128 GBGPU HBM3e | 4800 GB/s | yes | 371 tok/s4-bit quantization |
| DGX Station | 748 GBCoherent Memory | 7100 GB/s | yes | 433 tok/s4-bit quantization |
| ET900N G3 | 748 GBCoherent Memory | 7100 GB/s | yes | 433 tok/s4-bit quantization |
| DGX B200 | 1440 GBGPU HBM3e | 8000 GB/s | yes | 451 tok/s4-bit quantization |
| GB200 NVL72 | 13400 GBGPU HBM3e | 8000 GB/s | yes | 451 tok/s4-bit quantization |
| GB300 NVL72 | 20000 GBGPU HBM3e | 8000 GB/s | yes | 451 tok/s4-bit quantization |
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.
Demonstrates LiteLLM Auto Router with Pi Coding Agent, Ollama, and oMLX on Apple Silicon, comparing heuristic versus LLM classifiers for Qwen3.5 4B, Gemma 4, and Qwen3.8 27B.
Benchmarking Ornith 1.5 9B and 35B, Qwen 3.8 27B, and Gemma 4 on an RTX 5090 using llama-server with 4-bit weights and FP16 KV cache via code-graded metrics.
Explores running a 26B parameter MoE model on limited memory, comparing weight sizes across FP16/BF16, 8-bit, 4-bit, 3-bit, and 2-bit quantization levels.
Tests Qwen 3.8 27B, Muse Glimmer 30B and Gemma 4 26B on an RTX 5090 using Ollama, evaluating consistency across 12 accounting questions run 10 times each.
Side-by-side coding test using Claude Code on Ollama to generate Space Invaders, Breakout, and Tetris HTML canvas games with Qwen 3.8 27B, Muse Glimmer, and Gemma 4 on an RTX 5090.
Side-by-side coding test using Claude Code on Ollama to generate Space Invaders, Breakout, and Tetris HTML canvas games with Qwen 3.8 27B, Muse Glimmer, and Gemma 4 on an RTX 5090.
Regina test set results compare Gemma 26B and Qwen 35B using IQ4 quants from unsloth, identical sampling parameters, prompts, harness, and llama-server versions on Nvidia GPUs.
Qwen 3.6 35B A3B and Gemma 4 26B A4B are tested on updating documentation and verifying tests for mature legacy code using the swing-extras project.
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.
Deploys DiffusionGemma 26B-A4B in NVFP4 with vLLM on RTX PRO 6000, benchmarking its 256-token parallel generation against standard Gemma 4 to demonstrate a 6.73x speedup.
Tests compare Google gemma-4-26B-A4B-it-qat-q4_0-gguf against unsloth gemma-4-26B-A4B-it-GGUF using reasoning, agency, coding, and memory evaluations on a system with 16GB VRAM and 32GB DDR4 RAM.
Gemma 4 12B with QAT and QWEN 3.6 are tested as Coding Agents in VS Code on an M5 Max (128GB) using real-world coding tasks instead of benchmarks or synthetic tests.
Tests Gemma 4 26B A4B versus Qwen 3.6 35B A3B on a 16GB VRAM system using a six-test suite covering adherence, memory, implementation, architecture, reasoning, and agency.