{
"$type": "site.standard.document",
"bskyPostRef": {
"cid": "bafyreiaebh5otoxmbec7qr5e4ktvk2jly63hoac4ehms42ouk4ecmjg7wy",
"uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mkpsj2mn5qz2"
},
"path": "/t/squeeze-gemma-4-26b-on-a-4060ti-with-nvfp4/175654#post_2",
"publishedAt": "2026-04-30T13:01:02.000Z",
"site": "https://discuss.huggingface.co",
"tags": [
"CyberFitz/gemma-4-26B-A4B-it-NVFP4",
"google/gemma-4-26B-A4B-it",
"openclaw/openclaw"
],
"textContent": "Good question. Short version: **probably not “comfortably” on a 5060 Ti 16GB yet** , at least not in a clean plug-and-play OpenClaw setup.\n\nWhat matters:\n\n 1. **NVFP4 availability**\n\n\n * There are community NVFP4 checkpoints for Gemma 4 26B-A4B, but these are not the same as mainstream GGUF flows.\n * Example model card: CyberFitz/gemma-4-26B-A4B-it-NVFP4\n\n\n 2. **VRAM headroom**\n\n\n * Even that card reports about **~16 GB model size** and around **~18 GB minimum GPU memory** for serving, before comfortable KV cache headroom.\n * On a 16GB card, it may load only with tight limits / offloading and then feel slow.\n\n\n 3. **“No vision tower”**\n\n\n * Gemma 4 26B-A4B is a multimodal architecture; removing vision tower is not a standard toggle in typical runtimes.\n * You can run **text-only inference** without sending images, but physically stripping vision components is model surgery and usually breaks compatibility unless specifically supported.\n\n\n 4. **OpenClaw compatibility**\n\n\n * OpenClaw is the orchestration layer; real support depends on backend/runtime kernels (vLLM/TensorRT/llama.cpp/Ollama path you use).\n * If your backend doesn’t support this NVFP4 format end-to-end, it won’t help.\n\n\n\nPractical recommendation:\n\n * If you want reliability on 16GB today, use a **text-focused quantized path** with proven OpenClaw backend support.\n * If you want Gemma 4 26B NVFP4 specifically, expect experimentation and likely compromises (lower context, offload, slower throughput).\n\n\n\nBase model reference: google/gemma-4-26B-A4B-it\nOpenClaw repo/docs entry point: openclaw/openclaw",
"title": "Squeeze Gemma 4 26b on a 4060ti with NVFP4"
}