TL;DR: 12GB is the cheapest ticket into the 14B class. Qwen3-14B at Q4_K_M (9GB file, ~11.5GB VRAM at 8K context) is the smartest dense model that fits, at roughly 25–32 tok/s on an RTX 3060 12GB. Qwen3.5-9B is the faster daily driver with room for real context. The tier’s trap: a 14B at 8K context uses almost all of the card — one browser tab too many and it silently spills to CPU.

	Qwen3-14B Q4_K_M	Qwen3.5-9B Q4_K_M	Gemma 4 12B QAT Q4
Best for	Hardest tasks that fit: reasoning, code review	Coding agents, chat, vision, everyday work	Writing quality, instruction following
Download size	9GB	5.68GB	~7GB
Comfortable context on 12GB	8K (16K with KV quantization)	16K–32K	16K+
Speed on an RTX 3060 12GB	~25–32 tok/s	~38–42 tok/s	~30 tok/s class
The catch	Fills the card — the 8K-context ceiling is real	Same model the 8GB crowd runs	12B knowledge ceiling

Honest take: run Qwen3.5-9B as the default and pull Qwen3-14B for the questions the 9B fumbles. If you’re shopping rather than running what you own, the used RTX 3060 12GB at ~$289 is still the best dollars-per-VRAM deal in the entire GPU market — and that’s exactly why this tier is worth optimizing instead of upgrading.

The 12GB tier — used RTX 3060 12GB, RTX 4070, 4070 Super, RX 6700 XT — is the sweet spot of the used market in 2026. Our VRAM tier guide covers the whole ladder and the 8GB edition covers the rung below; this page is the dedicated 12GB list. Before committing to a download, you can sanity-check any model + context combination against your exact card with our VRAM calculator.

What the extra 4GB over an 8GB card actually buys is one of two things, and you have to pick: a smarter model (the 14B class at Q4 fits fully in VRAM — it doesn’t on 8GB) or more context for a mid-size model (the same Qwen3.5-9B that’s squeezed to 8K on an 8GB card runs 32K here). It usually does not buy both at once.

The real VRAM budget on a “12GB” card

Same arithmetic as every tier, different ceiling. Windows, the desktop compositor, and a hardware-accelerated browser hold several hundred MB before you load a model; plan on 11 to 11.5GB usable. Three things fit inside it:

1. Model weights — the GGUF file size, near enough
2. KV cache — grows linearly with context length
3. Compute buffers — a few hundred MB of scratch

A 14B model at Q4_K_M is a 9GB file, but at 8K context the whole load lands around 11.5GB per LLM Configurator’s measurements — which is why the 14B class is described as fitting a 12GB card rather than rattling around in one. The rule of thumb from LocalLLM.in’s 2026 VRAM guide holds: pick a quant file 1–2GB smaller than the card and give the rest to cache.

The picks, by what you actually do

The reason to own 12GB: Qwen3-14B

The 14B class is what this tier exists for, and Qwen3-14B is the pick. The Q4_K_M GGUF is 9GB (Q5_K_M 10.5GB — fits, but eats your context budget; Q6_K 12.1GB — doesn’t fit), and the 14B class runs fully in VRAM at roughly 25–32 tok/s on an RTX 3060 12GB per Local AI Master’s tier testing. Past reading speed, fully resident, no offload.

One naming note so you don’t chase a ghost: Qwen3.5 and 3.6 never shipped a dense 14B. The current Qwen line jumps from 9B to 27B (and the 27B needs a 24GB card). InsiderLLM’s mid-2026 Qwen roundup makes the same point — Qwen3-14B remains the current 14B, not a leftover. It’s the smartest dense model a 12GB card can hold, and it stays the pick until that changes.

ollama pull qwen3:14b
ollama run qwen3:14b
>>> /show info
  Model
    architecture        qwen3
    parameters          14.8B
    quantization        Q4_K_M

Same-size alternatives, same ~9GB Q4_K_M footprint, different personalities: Phi-4 14B for step-by-step reasoning (Hardwarepedia measures it at ~11.7GB total at 8K context, ~20–25 tok/s on the 3060 — the 3060 is literally the cheapest card that runs it fully in VRAM), and DeepSeek-R1 14B if you want visible chain-of-thought. Swapping between them costs nothing but disk.

Daily driver and coding: Qwen3.5-9B

The best all-rounder on 8GB is a better all-rounder on 12GB, because the 5.68GB Q4_K_M file leaves 5GB+ for cache — that’s 16K–32K context without KV-cache tricks. ModelFit’s RTX 3060 page puts it at ~38–42 tok/s, image input included. For Cline or Continue.dev wired to a local backend (setups on aicoderscope.com), the 9B at 32K context beats the 14B at 8K for agent work: agents die by context starvation more often than by model IQ. The honest-limits piece covers where any local model still loses to a frontier API.

If you want a dedicated code model instead of a generalist, Qwen2.5-Coder 14B still tops the HumanEval charts for this tier (Q4_K_M is 8.7GB, ~25–32 tok/s per Local AI Master) — but it’s a 2024-generation model with 2024-era tool-calling, so use it for autocomplete and FIM, not agents. Mistral’s Devstral Small 2 is the wrong answer here: the 24B needs ~15GB at Q4 and belongs to the 16GB tier.

Writing and instruction-following: Gemma 4 12B QAT

Google’s QAT checkpoint puts the 12B in ~7GB of VRAM per Unsloth’s Gemma 4 QAT table with near-original quality, since the quantization loss was trained out rather than bolted on. On an 8GB card it’s the “biggest thing that fits”; on 12GB it’s the comfortable pick — 16K+ context with headroom to spare. Prose and instruction-following are still noticeably better than anything else in its weight class. Background on the QAT release is in our Gemma 4 QAT guide.

What you cannot run on 12GB (stop trying)

The honest section. These are the models people force onto 12GB cards, and the specific wall each one hits:

Model	Q4-class footprint	Tier it actually needs
Gemma 4 26B-A4B QAT	~15GB	16GB
GPT-OSS 20B	~12.8GB weights, ~14GB loaded	16GB (details)
Qwen3.6-27B	~16.8GB	24GB (details)
Qwen3.6-35B-A3B	~22GB	24GB
Any dense 70B	~43GB	48GB, or heavy offload

GPT-OSS 20B is the painful one — at ~12.8GB for the Q4-class weights per WillItRunAI’s measurements, it almost fits, and “almost” means Ollama loads it split across GPU and CPU and you get single-digit tok/s while the fans tell you everything is fine. Gemma 4 26B-A4B QAT at ~15GB is the model most worth upgrading for; more on that below.

Partial offload is always available and almost never worth it at this tier: offloading a quarter of a 20B model costs you more speed than dropping to a fully-resident 14B. If you want the big models occasionally, rent instead — a RunPod A100 runs any of them full-speed for under $1.50/hour, and the rent-vs-buy math says occasional use never justifies a card upgrade.

The 8K-context trap: the problem you will actually hit

Here’s the failure mode specific to this tier, and the fix.

You pull Qwen3-14B, it runs beautifully at 25+ tok/s for a week. Then one day it generates at 6 tok/s and nothing you changed explains it. What happened: the 14B at 8K context wants ~11.5GB, your browser opened a WebGL tab that took 700MB of VRAM, and the pieces no longer fit. Two silent things follow:

• Ollama splits the model. ollama ps shows 52%/48% CPU/GPU in the PROCESSOR column instead of 100% GPU, and decode speed craters. Diagnosis walkthrough in Ollama not using GPU.
• Or the NVIDIA driver pages VRAM to system RAM — the Windows sysmem-fallback behavior, which ollama ps can’t even see. That one’s covered in the shared GPU memory fix.

The fix order on a 12GB card:

1. Quantize the KV cache first. OLLAMA_FLASH_ATTENTION=1 plus OLLAMA_KV_CACHE_TYPE=q8_0 — set on the service, not your shell — roughly halves cache VRAM at near-zero quality cost. On this tier that’s the difference between the 14B at 8K and the 14B at 16K. Full config in the tokens-per-second tuning guide.
2. Set context deliberately, and know who overrides whom. App settings and Modelfiles can silently raise num_ctx past what the card fits — the num_ctx priority order explains the precedence.
3. Close the GPU-hungry apps before blaming the model. On 12GB you have ~500MB of slack at 8K context with a 14B loaded. A browser with hardware acceleration can eat all of it.
4. If you need more room, drop the model, not the quant floor. Q4_K_M of the 9B with huge context beats Q3 of the 14B for almost every real workload — the quantization quality-loss numbers show the cliff sits at Q3.

Hard CUDA out of memory errors instead of silent slowdown? Different failure, own fix list.

Which 12GB card, and whether to buy one at all

Card	Price (Aug 2026)	Bandwidth	Verdict
RTX 3060 12GB used	~$289 avg, $278–$300 fair range	360 GB/s	The value king of the whole market. Cheapest full-VRAM 14B machine that exists
RTX 4070 used	~$485	504 GB/s	~40% faster tok/s, same model ceiling. Buy for speed, not capability
RX 6700 XT used	varies	384 GB/s	Fine with Vulkan; expect setup friction on ROCm

The uncomfortable truth about this tier: the RTX 4070 runs the same models as a $289 RTX 3060, just faster — bandwidth sets decode speed, but VRAM sets the ceiling, and both cards have the same ceiling. Roughly 25 vs 35 tok/s on a 14B (the 4070 figure is our bandwidth-scaled estimate — 504 vs 360 GB/s — not a measured number). Both are past reading speed. If you’re choosing between a used 4070 at $485 and a used 3060 at $289 for AI specifically, take the 3060 and bank the $200 toward the next tier.

If you want to run the next tier up

The first models worth upgrading for are Gemma 4 26B-A4B QAT (~15GB) and GPT-OSS 20B (~14GB loaded) — both need 16GB, and both are a real step up from the 14B class, not a side-step.

The card that unlocks them is the RTX 5060 Ti 16GB — ~$589 new / ~$460 used on eBay as of August 2026, up from its $429 MSRP thanks to the memory-price squeeze. We benchmarked what the jump actually unlocks in the 8GB vs 16GB comparison, and the full ladder logic lives in the GPU buying guide. If the budget stretches past $1,000, skip 16GB entirely — the used RTX 3090 at 24GB opens the Qwen3.6-27B and 35B-A3B class and remains the enthusiast default.

Not ready to buy? Rent the difference: a RunPod A100 at under $1.50/hour runs everything in the “cannot run” table above. At a few hours a week, that’s years of rental before a new card pays for itself.

Setup that works, start to finish

Tested against Ollama v0.32.6 (August 4, 2026):

# the smart one
ollama pull qwen3:14b

# the fast one
ollama pull qwen3.5:9b

# confirm full GPU residency after first run
ollama ps
NAME          ID          SIZE     PROCESSOR    UNTIL
qwen3:14b     f4e8a2c1    11 GB    100% GPU     4 minutes from now

Anything other than 100% GPU in the PROCESSOR column: apply the KV-cache quantization above, lower the context, or close whatever else is holding VRAM — in that order. For llama.cpp-based alternatives with finer offload control, aifoss.dev covers the open-source stack.

FAQ

Is 12GB enough for local AI in 2026, or is it obsolete? It’s the best value tier in the market. The 14B class fits fully in VRAM, the used 3060 costs ~$289, and nothing under $460 runs anything meaningfully smarter. “Obsolete” starts at 8GB, not 12.

Should I run the 14B at Q5_K_M since the file (10.5GB) technically fits? Only at 4K context, and it’s rarely worth it — the Q5-over-Q4 quality gain is small and the context you give up is not. Q4_K_M plus KV-cache quantization is the better trade on 12GB.

Can I run Gemma 4 26B-A4B on 12GB with offload? It loads, at single-digit tok/s with more than a quarter of the weights on CPU. If the 26B is what you want, that’s the signal to go 16GB — it’s the model that headlines the upgrade section above.

What about image generation on 12GB? Comfortable for SD 1.5/SDXL and workable for FLUX at FP8 — a genuinely better experience than 8GB. Start with the ComfyUI Windows setup guide.

RTX 3060 12GB vs RTX 4060 Ti 16GB used — which for AI? The 4060 Ti 16GB if the prices are close: VRAM ceiling beats bandwidth at equal money. But used 4060 Ti 16GB cards rarely sell near $289, and at a $150+ gap the 3060 wins on value again.

Recommended Gear

• RTX 3060 12GB — the best dollars-per-VRAM deal in the 2026 GPU market
• RTX 4070 — same 12GB ceiling, ~40% more speed for $200 more
• RTX 5060 Ti 16GB — the upgrade that unlocks the 26B QAT / GPT-OSS 20B tier

Sources

• Qwen3-14B-GGUF quant sizes (Q4_K_M 9GB, Q5_K_M 10.5GB, Q6_K 12.1GB) — Unsloth on Hugging Face
• Qwen3-14B: Specifications and GPU VRAM Requirements — APXML
• Best LLM for 12GB VRAM (2026) — Local AI Master
• Best Coding LLM for 12GB VRAM (2026) — Local AI Master
• RTX 3060 12GB Local LLM picks, ~38–42 tok/s — ModelFit
• Best Local Coding Models Ranked by VRAM Tier — LLM Configurator
• Best Qwen Models Ranked: Which to Run Locally (Mid-2026) — InsiderLLM
• Gemma 4 QAT VRAM table (12B ≈ 7GB, 26B-A4B ≈ 15GB) — Unsloth Documentation
• Phi-4 14B Hardware Requirements — Hardwarepedia
• DeepSeek-R1 14B: Specifications and GPU VRAM Requirements — APXML
• Qwen2.5-Coder-14B-Instruct-GGUF (Q4_K_M 8.7GB) — Bartowski on Hugging Face
• GPT-OSS 20B VRAM Requirements (12.8GB Q4_K_M) — WillItRunAI
• Qwen 3.6 27B VRAM Requirements (16.8GB Q4_K_M) — WillItRunAI
• RTX 3060 12GB Used GPU Price & Fair Asking Range — ResalePrices
• RTX 4070 Price Tracker US, Aug 2026 — Best Value GPU
• RTX 5060 Ti 16GB Price History & Tracker — GPU PRIX
• Ollama VRAM Requirements: Complete 2026 Guide — LocalLLM.in
• Ollama Release Notes, v0.32.6 — releases.sh

Last updated August 11, 2026. Prices and specs change; verify current rates before purchasing.