Running an AI model on your own PC is free and private, but the hardest part is choosing what to download. The right model depends mainly on how much video memory (VRAM) your graphics card has: a model that fits entirely in VRAM runs far faster than one that spills into system RAM. Since Ollama has just ended its free cloud allowance, picking a model that suits your own hardware matters more than ever.
Forge PC hasn't benchmarked these models. Our picks are based on the official model pages in Ollama's library, as checked on 17 November 2026, including each model's download sizes, context window, supported inputs and licence. We don't quote developers' own benchmark claims. Sizes are the Ollama download sizes; you'll need some extra memory on top for the context window, as our guide to LLM context length and VRAM explains.
Our picks at a glance
Pick | Model | Ollama download size | Context | Input | Suits |
|---|
Best all-rounder | Gemma 4 12B | 7.7-8.0GB | 256K | Text, images | 12GB GPUs and up |
Best for 8GB GPUs | Gemma 4 E4B | 6.6-9.5GB | 128K | Text, images | 8GB GPUs, laptops |
Best for coding on a budget | MiMo-V2.6 9B | 6.6GB (Q4_K_M) | 256K | Text, images | 8-12GB GPUs |
Best for reasoning on 16GB | gpt-oss 20B | 14GB | 128K | Text | 16GB GPUs |
Best for 24GB GPUs | Qwen3.8 27B | 18GB (Q4_K_M) | 256K | Text, images | 24GB GPUs |
Fastest big model | Gemma 4 26B (MoE) | 16-19GB | 256K | Text, images | 24GB GPUs |
Best tiny model | LFM3 1.7B | 1.1-1.7GB | 256K | Text | Any PC, background agents |
Best for 96GB+ machines | gpt-oss 120B | 65GB | 128K | Text | Ryzen AI Max and RTX Spark systems |
Best all-rounder: Gemma 4 12B
Gemma 4, from Google DeepMind, is described on Ollama's page as designed for reasoning, agentic workflows, coding and multimodal understanding, with configurable thinking modes and native function calling. The 12B model is the sweet spot for most gaming PCs: Ollama lists it at 7.7-8.0GB with a 256K-token context window, and it accepts images as well as text. On a 12GB or 16GB card it leaves room for a decent context. Our guide to running Gemma 4 locally covers setup.
Best for 8GB GPUs: Gemma 4 E4B
The "E" in E2B and E4B stands for "effective" parameters, and Google says these smaller models are made for edge devices such as laptops and phones. E4B has 4.5 billion effective parameters (8 billion including embeddings), a 128K context window and image input. Ollama lists the default download at 6.6-9.5GB depending on version, so pick the smaller quantisation on an 8GB card. If you're very tight on memory, E2B starts at 4.3GB. See how to run a local LLM on 8GB of VRAM.
Best for coding on a budget: MiMo-V2.6 9B
Xiaomi's MiMo-V2.6-Distill-Qwen-9B is a 9B model fine-tuned from Qwen3.5-9B for code and agent workflows. Ollama's page says it can understand unfamiliar code, investigate failures, draft fixes and use screenshots, with thinking and tool use and a 256K context. The Q4_K_M version is 6.6GB, and it's released under the MIT licence. It pairs well with coding tools; see running MiMo-V2.6 locally and using Continue with Ollama.
Best for reasoning on 16GB: gpt-oss 20B
OpenAI's open-weight gpt-oss models are built for reasoning and agentic tasks, with adjustable reasoning effort (low, medium or high) and the permissive Apache 2.0 licence. OpenAI quantises the mixture-of-experts weights to MXFP4, which it says lets the 20B model run on systems with as little as 16GB of memory. Ollama lists it at 14GB with a 128K context. It's text-only. See how to run gpt-oss locally.
Best for 24GB GPUs: Qwen3.8 27B
If you have an RTX 4090, RTX 5090 or another 24GB-plus card, Alibaba's Qwen3.8 27B is the step up. Ollama describes it as delivering gains across coding, professional work, research and long-horizon agentic tasks, with native image and video understanding. The standard Q4_K_M download is 18GB with a 256K context window; higher-precision Q8_0 is 30GB. Our Qwen3.8 guide has the details.
Fastest big model: Gemma 4 26B
Gemma 4 26B is a mixture-of-experts model with only 4 billion parameters active at a time, which is designed to make it quicker to run than a dense model of the same total size, though you still need memory for the full 16-19GB download. Our MoE vs dense LLM explainer covers the trade-off. If you prefer a dense model, Gemma 4 31B is 19-20GB.
Best tiny model: LFM3 1.7B
LFM3-1.7B is a text model for local agents and specialised applications. Ollama's page says it takes about 1GB of memory at 4-bit precision, so it fits on any phone or laptop, and it's post-trained for agent tools with a 256K context. Use it for quick background tasks rather than as a main chat model.
Best for 96GB+ machines: gpt-oss 120B
OpenAI says the larger gpt-oss model fits on a single 80GB GPU, and Ollama lists it at 65GB. That's out of reach for a normal graphics card, but systems with large pools of unified memory, such as Ryzen AI Max PRO 495 mini PCs with 192GB, can hold it. See our best PC for local AI guide.
How to choose
Match the model to your VRAM. Leave a couple of gigabytes free for context. Our best GPU for local LLMs guide covers upgrades.
Pick the right quantisation. Q4_K_M is the usual balance of size and quality; see GGUF quantisation explained.
Check the licence if you'll use a model for work. Ollama's pages list Apache 2.0 for gpt-oss and MIT for MiMo-V2.6.
Start simple. Ollama or LM Studio will download and run all of these; see how to run AI models locally.
Our verdict
For most gaming PCs with a 12GB or 16GB graphics card, Gemma 4 12B is the best place to start: small enough to fit with room for context, with image input and a 256K window. Choose gpt-oss 20B for reasoning on a 16GB card, MiMo-V2.6 9B for coding on a budget, and Qwen3.8 27B or Gemma 4 26B if you have 24GB of VRAM. On an 8GB card, start with Gemma 4 E4B. Forge PC hasn't benchmarked these models; picks are based on official model information.