Unverified50% confidenceFactExact time
Gemma 3 12B官方支持的上下文长度为128K token，但在Ollama中默认num_ctx仅为2048
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/13/2026
First Seen
Valid until: 10/11/2026
Sources
本地Gemma 3 12B变身AI Agent实战指南：工具调用与框架选择
redditr/aiagents7/12/2026
Related Claims
UnverifiedGemma 3 12B的FP16版本约需24GB显存，Q4_K_M版本降至约7-8GB，推理速度提升约2-3倍，结构化输出准确率仅损失约2-5%72% similarUnverifiedGemma 4 26B模型约17GB，适合24GB以上显存的显卡，编程效果最佳69% similarUnverified2025年发布的 Gemma 3 系列扩展了多模态能力，支持图像理解，并将上下文窗口扩展至128K tokens67% similarUnverifiedRTX 5090上Qwen3.6 35B-A3B推理速度达220 token/s，显存占用约23GB65% similarUnverified最新模型的上下文长度已达128K token64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/501966API
curl https://kongchang.com/api/v1/knowledge/claims/501966MCP
get_claim(id=501966)