What LLMs Fit in 12 GB of VRAM? (Oct 2026)

GeForce RTX 3060 12GB (desktop) is the catalog SKU in this 12 GB tier with the smallest net budget: 11.00 GiB after the sourced reserve. 43 catalog models fit at a file-backed artifact. The month follows dataset 2026-10-02.1 (Oct 2026), not the build clock.

Catalog dataset 2026-10-02.1 (Oct 2026).

Catalog SKUs in this tier

SKU	Marketed capacity	Net budget
GeForce RTX 3060 12GB (desktop) Smallest net budget	12 GB GDDR6	11.00 GiB
GeForce RTX 3080 12GB (desktop)	12 GB GDDR6X	11.00 GiB
GeForce RTX 4070 12GB (desktop, GDDR6X)	12 GB GDDR6X	11.00 GiB
GeForce RTX 4070 SUPER 12GB (desktop)	12 GB GDDR6X	11.00 GiB
GeForce RTX 5070 12GB (desktop)	12 GB GDDR7	11.00 GiB

Catalog models that fit

Each row is one file-backed artifact on the representative SKU. A missing quantization is omitted, not replaced. Context search stops at the sourced cap. With no sourced cap, the only verdict shown is at 8192 tokens.

When catalog bandwidth is missing, the decode ceiling says not available. That is not a memory verdict.

Model	Artifact	Max context that fits	Max context including tight	Verdict at 8192 tokens	Decode ceiling	Links
Qwen2.5-0.5B-Instruct	GGUF Q4_K_M (Qwen) `gguf-q4_k_m` · `llama.cpp`	32,768	32,768	8192 Runs comfortably 0.68 GiB–1.07 GiB	Not available
Qwen3-0.6B	GGUF Q4_K_M (Unsloth) `gguf-q4_k_m` · `llama.cpp`	40,960	40,960	8192 Runs comfortably 1.37 GiB–1.81 GiB	Not available
Gemma 3 1B IT	GGUF Q4_0 (Google QAT) `gguf-q4_0` · `llama.cpp`	32,768	32,768	8192 Runs comfortably 1.10 GiB–1.50 GiB	Not available
Llama 3.2 1B Instruct	Official BF16 safetensors `bf16` · `vllm`	131,072	131,072	8192 Runs comfortably 3.18 GiB–5.06 GiB	Not available
Qwen2.5-1.5B-Instruct	GGUF Q4_K_M (Qwen) `gguf-q4_k_m` · `llama.cpp`	32,768	32,768	8192 Runs comfortably 1.38 GiB–1.80 GiB	Not available
Qwen2.5-Coder-1.5B-Instruct	GGUF Q4_K_M (Qwen) `gguf-q4_k_m` · `llama.cpp`	32,768	32,768	8192 Runs comfortably 1.38 GiB–1.80 GiB	Not available
SmolLM2-1.7B-Instruct	GGUF Q4_K_M (Hugging Face) `gguf-q4_k_m` · `llama.cpp`	8,192	8,192	8192 Runs comfortably 2.61 GiB–3.11 GiB	Not available
Qwen3-1.7B-Base	GGUF Q4_K_M (mradermacher) `gguf-q4_k_m` · `llama.cpp`	32,768	32,768	8192 Runs comfortably 2.03 GiB–2.49 GiB	Not available
Qwen3-1.7B	GGUF Q4_K_M (Unsloth) `gguf-q4_k_m` · `llama.cpp`	40,960	40,960	8192 Runs comfortably 2.03 GiB–2.49 GiB	Not available
Qwen2.5-3B-Instruct	GGUF Q4_K_M `gguf-q4_k_m` · `llama.cpp`	32,768	32,768	8192 Runs comfortably 2.37 GiB–2.80 GiB	Not available
Qwen2.5-Coder-3B-Instruct	GGUF Q4_K_M (Qwen) `gguf-q4_k_m` · `llama.cpp`	32,768	32,768	8192 Runs comfortably 2.37 GiB–2.80 GiB	Not available
Llama 3.2 3B Instruct	Official BF16 safetensors `bf16` · `vllm`	12,868	41,107	8192 Fits 7.48 GiB–9.39 GiB	Not available
Qwen2.5-VL-3B-Instruct	GGUF Q4_K_M (Unsloth) `gguf-q4_k_m` · `llama.cpp`	128,000	128,000	8192 Runs comfortably 2.20 GiB–2.63 GiB	Not available
Phi-3.5-mini-instruct	GGUF Q4_K_M (bartowski) `gguf-q4_k_m` · `llama.cpp`	18,187	23,609	8192 Runs comfortably 5.35 GiB–5.98 GiB	Not available
Phi-4-mini-instruct	GGUF Q4_K_M (Unsloth) `gguf-q4_k_m` · `llama.cpp`	53,843	70,076	8192 Runs comfortably 3.45 GiB–3.94 GiB	Not available
Qwen3-4B	GGUF Q4_K_M `gguf-q4_k_m` · `llama.cpp`	40,960	40,960	8192 Runs comfortably 3.58 GiB–4.08 GiB	Not available
Qwen3-4B-Instruct-2507	GGUF Q4_K_M (Unsloth) `gguf-q4_k_m` · `llama.cpp`	47,826	62,253	8192 Runs comfortably 3.58 GiB–4.08 GiB	Not available
Gemma 3 4B IT	GGUF Q4_0 (Google QAT) `gguf-q4_0` · `llama.cpp`	131,072	131,072	8192 Runs comfortably 3.33 GiB–3.79 GiB	Not available
Qwen3-VL-4B-Instruct	GGUF Q4_K_M (Qwen) `gguf-q4_k_m` · `llama.cpp`	47,826	62,253	8192 Runs comfortably 3.58 GiB–4.08 GiB	Not available
Mistral-7B-Instruct-v0.3	GGUF Q4_K_M (bartowski) `gguf-q4_k_m` · `llama.cpp`	32,768	32,768	8192 Runs comfortably 5.20 GiB–5.72 GiB	Not available
DeepSeek-R1-Distill-Qwen-7B	GGUF Q4_K_M (Unsloth) `gguf-q4_k_m` · `llama.cpp`	86,646	121,963	8192 Runs comfortably 4.92 GiB–5.42 GiB	Not available
OpenMath-Nemotron-7B	GGUF Q4_K_M (mradermacher) `gguf-q4_k_m` · `llama.cpp`	86,646	121,963	8192 Runs comfortably 4.92 GiB–5.42 GiB	Not available
Qwen2-7B-Instruct	GGUF Q4_K_M (Qwen) `gguf-q4_k_m` · `llama.cpp`	32,768	32,768	8192 Runs comfortably 4.92 GiB–5.42 GiB	Not available
Qwen2.5-7B-Instruct	GGUF Q4_K_M `gguf-q4_k_m` · `llama.cpp`	32,768	32,768	8192 Runs comfortably 4.92 GiB–5.42 GiB	Not available
Qwen2.5-Coder-7B	GGUF Q4_K_M (mradermacher) `gguf-q4_k_m` · `llama.cpp`	32,768	32,768	8192 Runs comfortably 4.92 GiB–5.42 GiB	Not available
Qwen2.5-Coder-7B-Instruct	GGUF Q4_K_M (Qwen) `gguf-q4_k_m` · `llama.cpp`	32,768	32,768	8192 Runs comfortably 4.92 GiB–5.42 GiB	Not available
Qwen2.5-Math-7B-Instruct	GGUF Q4_K_M (bartowski) `gguf-q4_k_m` · `llama.cpp`	4,096	4,096	8192 Runs comfortably · Rough estimate 4.92 GiB–6.23 GiB	Not available
DeepSeek-R1-Distill-Llama-8B	GGUF Q4_K_M (Unsloth) `gguf-q4_k_m` · `llama.cpp`	36,179	51,545	8192 Runs comfortably 5.71 GiB–6.24 GiB	Not available
granite-3.3-8b-instruct	GGUF Q4_K_M (IBM) `gguf-q4_k_m` · `llama.cpp`	28,814	41,101	8192 Runs comfortably 5.98 GiB–6.53 GiB	Not available
Qwen3-8B	GGUF Q4_K_M `gguf-q4_k_m` · `llama.cpp`	31,467	40,960	8192 Runs comfortably 5.93 GiB–6.48 GiB	Not available	Qwen3-8B · GeForce RTX 3060 12GB (desktop) Qwen3-8B · GeForce RTX 4070 12GB (desktop, GDDR6X)
Qwen3-8B-Base	GGUF Q4_K_M (mradermacher) `gguf-q4_k_m` · `llama.cpp`	31,467	32,768	8192 Runs comfortably 5.93 GiB–6.48 GiB	Not available	Qwen3-8B-Base · GeForce RTX 4070 12GB (desktop, GDDR6X)
Qwen2.5-VL-7B-Instruct	GGUF Q4_K_M (Unsloth) `gguf-q4_k_m` · `llama.cpp`	86,646	121,963	8192 Runs comfortably 4.92 GiB–5.42 GiB	Not available
Qwen3-VL-8B-Instruct	GGUF Q4_K_M (Qwen) `gguf-q4_k_m` · `llama.cpp`	31,467	45,092	8192 Runs comfortably 5.93 GiB–6.48 GiB	Not available	Qwen3-VL-8B-Instruct · GeForce RTX 4070 12GB (desktop, GDDR6X)
GLM-4-9B-0414	GGUF Q4_K_M (Unsloth) `gguf-q4_k_m` · `llama.cpp`	32,768	32,768	8192 Runs comfortably 6.18 GiB–6.69 GiB	Not available	GLM-4-9B-0414 · GeForce RTX 3060 12GB (desktop) GLM-4-9B-0414 · GeForce RTX 4070 12GB (desktop, GDDR6X)
Gemma 3 12B IT	GGUF Q4_0 (Google QAT) `gguf-q4_0` · `llama.cpp`	21,364	49,849	8192 Fits 8.46 GiB–9.04 GiB	Not available
Mistral-Nemo-Instruct-2407	GGUF Q4_K_M (Unsloth) `gguf-q4_k_m` · `llama.cpp`	14,069	25,633	8192 Fits 8.34 GiB–8.94 GiB	Not available	Mistral-Nemo-Instruct-2407 · GeForce RTX 3060 12GB (desktop) Mistral-Nemo-Instruct-2407 · GeForce RTX 4070 12GB (desktop, GDDR6X)
phi-4	GGUF Q4_K_M (Unsloth) `gguf-q4_k_m` · `llama.cpp`	4,907	13,606	8192 Tight fit 9.97 GiB–10.62 GiB	Not available
Phi-4-reasoning	GGUF Q4_K_M (Unsloth) `gguf-q4_k_m` · `llama.cpp`	4,112	12,811	8192 Tight fit 10.12 GiB–10.77 GiB	Not available
Qwen3-14B	GGUF Q4_K_M `gguf-q4_k_m` · `llama.cpp`	5,453	16,328	8192 Tight fit 9.76 GiB–10.39 GiB	Not available	Qwen3-14B · GeForce RTX 3060 12GB (desktop) Qwen3-14B · GeForce RTX 4070 12GB (desktop, GDDR6X)
OpenMath-Nemotron-14B	GGUF Q4_K_M (mradermacher) `gguf-q4_k_m` · `llama.cpp`	4,614	13,676	8192 Tight fit 10.00 GiB–10.64 GiB	Not available
Qwen2.5-14B-Instruct	GGUF Q4_K_M `gguf-q4_k_m` · `llama.cpp`	4,614	13,676	8192 Tight fit 10.00 GiB–10.64 GiB	Not available
Qwen2.5-Coder-14B-Instruct	GGUF Q4_K_M (Qwen) `gguf-q4_k_m` · `llama.cpp`	4,614	13,676	8192 Tight fit 10.00 GiB–10.64 GiB	Not available
Qwen3.8-27B	GGUF UD-IQ1_S (Unsloth) `gguf-ud-iq1_s` · `llama.cpp`	31,666	81,266	8192 Fits · Rough estimate 6.54 GiB–8.14 GiB	Not available

Allowlisted combos

Mistral-Small-24B-Instruct-2501 · GeForce RTX 3060 12GB (desktop)