Best Local Vision Models by VRAM

Open-weight vision-language models sized by VRAM: Qwen3-VL 2B to 235B-A22B, Mistral Small 3.2, Gemma 3, Pixtral, and where each tier breaks.

The maintained version of this guide is Best Local Vision Models: Every GPU Tier. This page is kept so existing links keep working.

A local vision-language model lets a reader drop a screenshot into a chat window without the frame ever leaving the machine, but only if the model is ungated, fits the card, and answers the question actually being asked. The split between Apache 2.0 and gated licences is the practical filter most readers trip on first. Every figure below was pulled from the Hugging Face model API or the live Ollama tag pages, first on 2026-10-06 and re-read on 2026-11-16; the Ollama tag sizes, licences, gating flags and parameter counts all still matched on that date.

Quick reference

JobModelOllama tag (size)LicenceGatedContext
Sub-2 GB on-device OCR/chatQwen3-VL-2B-Instructqwen3-vl:2b-instruct (1.9 GB)Apache 2.0No256K (expandable to 1M)
4 GB tier, OCR + short captioningQwen3-VL-4B-Instructqwen3-vl:4b-instruct (3.3 GB)Apache 2.0No256K (expandable to 1M)
8 GB sweet spot, document + UIQwen3-VL-8B-Instructqwen3-vl:8b-instruct (6.2 GB)Apache 2.0No256K (expandable to 1M)
12 GB tier, images + documentsGemma 3 12Bgemma3:12b (8.2 GB)GemmaYes128K
16 GB tier, tight fit, no firm pickMistral Small 3.2 24Bmistral-small3.2:24b (15 GB)Apache 2.0No128K
24 GB tier, largest dense Qwen3-VLQwen3-VL-32B-Instructqwen3-vl:32b-instruct (21 GB)Apache 2.0No256K (expandable to 1M)
8 GB tier, gated alternativeLlama 3.2 Vision 11Bllama3.2-vision:11b (7.8 GB)Llama 3.2Yes128K
64 GB+ tier, MoE flagshipQwen3-VL-235B-A22B-Instructqwen3-vl:235b-a22b-instruct (143 GB)Apache 2.0No256K (expandable to 1M)
64 GB+ tier, gatedLlama 3.2 Vision 90Bllama3.2-vision:90b (55 GB)Llama 3.2Yes128K

Three notes on the tag and size columns. On Ollama, the plain qwen3-vl:2b, :4b, :8b, :32b and :235b tags resolve to the Thinking variants (the same digests as the -thinking tags, checked 2026-11-16); the Instruct models in this table are the -instruct tags, at the same sizes. The safetensors.total returned by the Hugging Face API for Qwen3-VL-32B is a parameter count, 33.4 billion, which is about 66.7 GB (62.1 GiB) at BF16’s two bytes per parameter; the 21 GB Ollama tag is a Q4_K_M quantisation of the 32B weights (the BF16 tag is listed at 67 GB). The cluster convention is to publish Ollama tag sizes, since that is what readers pull, and to fall back to the HF safetensors total when no Ollama tag is available. The Gemma 3 numbers come from the same Ollama library page, read on 2026-11-16.

Two jobs, two architectures

OCR and short captioning. The model receives a document, screenshot, or sign and returns the text or a one-line description. The hardest sub-task is reading dense tables and handwritten text. Qwen3-VL at 2B and 4B covers this band without licence friction: ungated, Apache 2.0, and under 3.5 GB as Ollama tags.

Visual reasoning over multi-image or video input. The model receives several images, a video, or a long document with figures and is asked to compare, locate, or summarise across them. The native context window matters: Qwen3-VL-8B and Qwen3-VL-32B both ship 256K native context, which the cards say is expandable to 1M, and the Qwen3-VL-32B card names “hours-long video with full recall and second-level indexing” as a documented capability. Gemma 3 12B and 27B ship a 128K context, enough for most document jobs, but Google’s model card lists text and images as inputs, not video.

The RAG setups page covers pairing a vision model with a text embedder for retrieval over a personal screenshot library; the how-much-VRAM guide covers the runtime overhead the table above does not show.

Pick by tier

Every use case at one memory budget, in one place, for the tiers not linked below: 6 GB | 12 GB | 24 GB | 32 GB.

Under 2 GB. qwen3-vl:2b-instruct at 1.9 GB is the smallest model in this table and the one to try on low-memory machines and older integrated-GPU laptops. The card is ungated and the Apache 2.0 licence removes the friction that Llama 3.2 Vision’s gate creates. Expect a clear step down on dense-document OCR versus the 4B and 8B variants.

3 to 4 GB. qwen3-vl:4b-instruct at 3.3 GB is the sub-4 GB choice for screenshots, single-page documents, and short UI captures. The earlier qwen2.5vl:3b (3.2 GB) is still on Ollama, but its Ollama context is 125K against Qwen3-VL’s 256K, and its Hugging Face card carries the Qwen Research licence rather than Apache 2.0.

6 to 9 GB. qwen3-vl:8b-instruct at 6.2 GB is the 8 GB sweet spot: ungated, Apache 2.0, 256K native context, and documented video understanding. gemma3:12b (8.2 GB) lands in the same VRAM band but is gated on Hugging Face; the Ollama tag resolves, but a reader loading from hf.co/google/gemma-3-12b-pt will hit a manual access request. The 8 GB tier guide picks Qwen3.5-9B instead, which also reads images: the Unsloth Q4_K_M weights plus the mmproj-F16 projector are 6,598,688,544 bytes (6.60 GB, 6.15 GiB), and Alibaba’s published comparison sets it against Qwen3-VL-30B-A3B rather than the 8B, so no same-size result decides between the two.

12 to 16 GB. Three routes diverge. mistral-small3.2:24b at 15 GB is the closest ungated fit in this band, not a firm pick: Apache 2.0, 128K context, and Mistral’s model card documents image input, but the Unsloth Q4_K_M weights plus the mmproj-F16 projector are 15,211,976,416 bytes (15.21 GB, 14.17 GiB), which does not fit a 12 GB card and leaves about 1.8 GiB of a 16 GiB card for the KV cache, image tokens and runtime overhead. The 16 GB tier guide makes no vision pick at that tier, with the same thin margin among its reasons. qwen2.5vl:32b at 21 GB is the previous-generation dense Qwen VLM, still on Ollama, but it needs the next tier up. llama3.2-vision:11b at 7.8 GB is the gated Llama alternative; the Ollama tag skips the Hugging Face access form, though the Llama 3.2 licence still applies.

20 to 24 GB. qwen3-vl:32b-instruct at 21 GB is the largest dense Qwen3-VL model: Apache 2.0, ungated, 33.4B parameters (about 66.7 GB at BF16, 21 GB as the Q4_K_M tag), 256K context, and the model card documents timestamp-grounded event localization for video. The 30B-A3B MoE variant (qwen3-vl:30b-a3b-instruct at 20 GB) sits in the same VRAM band with a different activation pattern; the Ollama tag is live. Qwen3-VL is no longer the newest Qwen vision line: the later dense Qwen3.5-27B, Qwen3.6-27B and Qwen3.8-27B also take image input, and Alibaba’s Qwen3.5-27B card reports it ahead of Qwen3-VL-235B-A22B on most vision rows, while the 235B stays ahead on six, including CC-OCR (81.5 against 81.0) and SimpleVQA (61.3 against 56.0).

64 GB+. qwen3-vl:235b-a22b-instruct at 143 GB Q4_K_M is the open-weight MoE flagship: 235B total / 22B active, Apache 2.0, ungated, 256K native context. The card documents the same spatial and video capabilities as the 32B variant; the practical difference is scale and the far heavier VRAM (or disk + offload) footprint. Llama 3.2 Vision 90B at 55 GB is the gated competitor at this scale.

Licences, where the traps are

The vision-model family splits cleanly into two camps. The Apache 2.0 camp covers Qwen3-VL (all six sizes), Qwen2.5-VL 7B and 32B (the 3B card carries the Qwen Research licence and the 72B card the Qwen licence), Mistral Small 3.2 24B, Pixtral 12B, and LLaVA-OneVision Qwen2-7B. Each can be deployed commercially without contacting the rights holder.

The gated camp covers Llama 3.2 Vision (Llama 3.2 Community License) and Gemma 3 (Gemma Terms of Use, with separate restrictions for derivative works). Llama 3.2’s community license requires a separate licence from Meta when a licensee’s products had more than 700 million monthly active users on the release date; Gemma’s terms add an explicit responsible-AI prohibited-use clause. Ollama serves both, but the underlying HF repo is gated, which means a fresh deployment that bypasses Ollama must complete the access form first.

Two specific cases to flag. The Qwen2.5-VL-3B-Instruct card lists the Qwen Research licence and the Qwen2.5-VL-72B-Instruct card the Qwen licence, rather than Apache 2.0; the open-model licences guide covers what that means in practice. Pixtral 12B is Apache 2.0 per the HF API but Ollama has no library page for it; confirm via the Mistral Hugging Face hub that the 12B 2409 build is the version a deployment needs.

How to actually run one

Two flags matter most for vision models. The Qwen3-VL family uses dynamic resolution with configurable min/max pixel bounds (the 32B preprocessor config sets 65,536 and 16,777,216 pixels as the lower and upper bounds, so the token cost of an image tracks its resolution). The Gemma 3 family uses a fixed input resolution: Google’s model card says images are “normalized to 896 x 896 resolution and encoded to 256 tokens each”. The Ollama gemma3 tag pulls without the Hugging Face access form, but the Gemma terms still apply.

The GGUF, AWQ, GPTQ and MLX comparison covers which format to pull. The what-quantisation costs you page covers how quality loss shows up on vision tasks; it is more visible than on chat because small alignment errors compound across multi-image prompts. The household-serving guide covers the network plumbing for putting a local VLM behind Open WebUI on the home LAN.

The bottom line

For most consumer cards, qwen3-vl:8b-instruct (6.2 GB, Apache 2.0, ungated, 256K context) is the Qwen3-VL option at the 8 GB sweet spot; it handles document OCR, screenshot Q&A, and short video clips without the friction of a gated repo. On a 16 GB card there is no firm pick: mistral-small3.2:24b (15 GB, Apache 2.0, 128K context) is the closest ungated fit, with under 2 GiB left for context and images; on a 24 GB card, qwen3-vl:32b-instruct (21 GB, Apache 2.0, 256K context) is the largest dense Qwen3-VL model. The two gated picks worth knowing exist: gemma3:27b at 17 GB for readers who already have access, and llama3.2-vision:90b at 55 GB for readers who accept the Llama 3.2 licence. Above 64 GB, qwen3-vl:235b-a22b-instruct (143 GB, Apache 2.0, 235B/22B MoE) is the largest open-weight pick here for hours-long video and long-context document understanding.

Two traps to avoid. First, do not pull a vision model from a gated Hugging Face repo without confirming the access form has been accepted; Ollama can bypass the gate, but the underlying licence still applies on every deploy. Second, do not assume a closed-weight benchmark number carries over to an open-weight model at the same parameter count; the open-weight cards publish their own numbers, and the leading open-weight models page is the cluster index for what is currently shipping.

Other GPUs and tasks: Local AI by VRAM · RSS