Best local AI models for NVIDIA 840M

The largest models that fit fully

The 30 largest of the 56 models that fit; every smaller model in the catalog fits too. Best quant means the highest quality compression whose weights and 4k context both sit inside the memory.

Close, but only with CPU offload

These need more than the card holds at their smallest practical quant, so part of the model runs from system memory (figures assume 32 GB of it). They work, several times slower.

How to read this

The NVIDIA 840M is an entry level laptop graphics card equipped with 2 GB of DDR3 video memory. This 2 GB limit dictates the maximum size of the artificial intelligence models you can run entirely on the hardware. To fit local models inside this tight memory space, you must use quantized versions. Quantization reduces the precision of the model weights to save space with minimal loss in output quality.

The best quantization level for your hardware depends on the size of the model. For the largest models that fit completely in your video memory, you must use highly compressed quants like Q4_K_M. This includes the 2.8B Allegro, the 2.7B Open-Sora Plan, the 2.6B LFM2, and the 2.6B Playground v2.5. These models utilize 1.9 GB to 2 GB of your video memory. Other models in this range include the 2.5B Stable Diffusion 3.5 Medium, the 2.5B Canary, and the 2.3B SeamlessM4T v2.

Medium sized models between 1.6B and 2.2B parameters can run at slightly better quantization levels such as Q5_K_M or Q6_K. The 2.2B Parler-TTS and the 2.2B Kimi K3 DSpark run well at Q5_K_M. The 2B SmolVLM, the 2B Stable Diffusion 3 Medium, the 2B Pyramid Flow, and the 2B Wav2Vec2 XLS-R fit at Q6_K. You can also run the 1.9B Moondream 2, the 1.7B Qwen3, and the 1.7B SmolLM2 at Q6_K while staying under your 2 GB limit.

Smaller models under 1.6B parameters can run at the high quality Q8_0 quantization level. This group includes the 1.6B StableLM 2, the 1.6B Sana, the 1.6B Zonos 0.1, and the 1.6B Dia. It also includes the 1.55B Whisper Large v3, the 1.5B ControlNet, the 1.5B Hunyuan-DiT, the 1.5B Stable Video Diffusion, and the 1.5B Whisper Large v2. Audio models like the 1.5B AudioGen, the 1.5B AudioLDM 2, and the 1.4B Tango 2 also run at Q8_0. The 1.1B TinyLlama and the 1.1B SantaCoder run at Q8_0 while using only 1.4 GB of video memory.

If you want to run larger models, you must use CPU offloading. This process splits the model between your video memory and your system RAM. For these setups, we assume you have 32 GB of system RAM. Offloading allows you to run the 3B SmolLM3, the 3B Replit Code v1.5, the 3B Kandinsky 3.1, the 3B Voxtral Mini, the 3B Orpheus TTS, and the 3B Higgs Audio v2. These models require 2.2 GB of video memory at Q4_K_M and 4.2 GB of system RAM.

Larger offloaded models include the 3.3B MusicGen, which needs 2.4 GB of video memory at Q4_K_M and 4.4 GB of system RAM. The 3.5B SDXL Turbo and the 3.5B SDXL Lightning need 2.6 GB of video memory at Q4_K_M and 4.6 GB of system RAM. The 3.417B Stable Diffusion XL needs 4.1 GB at FP8 and 6.1 GB of system RAM. Note that offloading slows down processing speeds significantly because system RAM is much slower than video memory.

When running text models, you must also consider the context window. Running a model at a standard 4k context window requires extra video memory to store the active conversation history. This extra memory usage can push a model over your 2 GB limit and trigger slow system RAM offloading. To prevent this slowdown, you should reduce your active context window size when your video memory is nearly full.