Here's what the current (2026) consensus says about the best local LLM for an RTX 3080 Ti (12 GB VRAM, 912 GB/s bandwidth): ## 🏆 Overall Best: Qwen 3 14B (Q4_K_M) This is the top pick across multiple ranking sites (BestLLMfor, llmhardware.io, hardware-corner). It's a 14B dense model from Alibaba, Apache 2.0 licensed, with a hybrid thinking/reasoning mode. - **VRAM at Q4:** ~9 GB (fits comfortably in 12 GB) - **Context:** 128k - **Speed:** ~52–63 t/s generation, ~1,774 t/s prompt processing - Why: matches Qwen 2.5 32B Base on STEM/code while being small enough to fit your card. The 12 GB VRAM is a hard ceiling at 14B — this is the sweet spot. ## 💻 Best for Coding: Qwen 3 Coder 14B (Q4_K_M) If your main use is coding, this is the top pick for 12 GB cards (PromptQuorum). Highest HumanEval among 14B models. - **VRAM at Q4:** ~9–10 GB - **Speed:** ~18–22 t/s on the 3080 Ti (thanks to its high bandwidth) - Set `--num-ctx 8192` minimum for coding, and enable Flash Attention to free up VRAM. - Strong alternative: **DeepSeek Coder 14B** (80+ languages). ## 🧠 Best for Reasoning/Math: Phi-4 Reasoning 14B or DeepSeek R1 Distill Qwen 14B - **Phi-4 Reasoning 14B** (MIT) — beats R1-Distill-Llama-70B on AIME/GPQA with far fewer params. ~9 GB VRAM, 32k context. - **DeepSeek R1 Distill Qwen 14B** (MIT) — AIME24 69.7, MATH-500 93.9. ~9 GB VRAM, 128k context. ## Other solid options that fit - **Mistral Nemo 12B Instruct** (Apache 2.0) — 128k context, strong multilingual. ~7 GB VRAM. - **Gemma 3 12B** (Google) — multimodal, 128k context, 140 languages. ~7 GB VRAM. - **Phi-4 14B** (MIT) — strong STEM bias. ~9 GB VRAM. ## Key practical notes for your 3080 Ti - **12 GB is the hard ceiling** — you're limited to 14B models at Q4 quantization. Anything larger requires offloading to system RAM, which tanks speed. - **Use Q4_K_M** as your starting quantization — it preserves ~85% of quality while fitting in VRAM. Drop to Q3_K_M only if a model barely fits. - **Context limits:** ~16k on Qwen3 14B, up to 32k on Qwen3 8B. You can't do 64k+ context on 14B models. - **Recommended tools:** Ollama (`ollama run qwen3:14b`) or LM Studio — both are one-command setups. - The card is genuinely fast for its class (912 GB/s bandwidth), so 14B models run at ~50–70 t/s — very usable. **Bottom line:** For general use, go with **Qwen 3 14B (Q4_K_M)**. For coding specifically, **Qwen 3 Coder 14B**. Both fit your 12 GB VRAM comfortably and give excellent speed on the 3080 Ti.