NVIDIA · 8 GB
The mainstream 4B–8B tier. Some 12B INT4 builds fit tightly with modest context.
Qwen3 8B
清單中排最前的模型可能是勉強適配或需要卸載的。日常使用建議選第一個標為 餘量充足 的結果，並保持上下文適中。
Ollama
One-command local chat and app integration. GPU and driver support varies by operating system. Confirm the current hardware matrix before buying hardware. Since v0.34.2 (2026-09-15) Ollama's first-run setup offered account sign-in, and the release notes said you could continue locally; since v0.40.1 (2026-10-07) the CLI Welcome screen completes onboarding on Enter and the intermediate account step is gone, while cloud models keep the sign-in flow. Since v0.40.0 (2026-09-25) Ollama runs MLX-supported model architectures on Apple silicon through MLX by default rather than its own runner; NVIDIA and CPU paths are unchanged.
開啟指南代表性開放權重模型
Qwen3 8B
- 估算記憶體
- ~6.8 GB
- 上下文
- 32K+
Qwen3 4B
- 估算記憶體
- ~3.6 GB
- 上下文
- 32K+
Gemma 3 4B
- 估算記憶體
- ~4.2 GB
- 上下文
- 128K
Llama 3.2 3B
- 估算記憶體
- ~2.8 GB
- 上下文
- 128K
Gemma 3 1B
- 估算記憶體
- ~1.4 GB
- 上下文
- 32K
適配以估算的總模型記憶體為準。驅動支援、KV 快取、多模態投影層、併發以及桌面應用都可能進一步壓縮可用餘量。目前適配：Qwen3 8B (勉強適配), Qwen3 4B (餘量充足), Gemma 3 4B (餘量充足), Llama 3.2 3B (餘量充足), Gemma 3 1B (餘量充足)。