Local AI by VRAM Tier - 8GB | 12GB | 16GB | 24GB | 32GB
Deep dives: Chat | Coding | Translation | Vision | Speech | Agents
Updated July 2026: the open vision-language field moved fast. Qwen3-VL now spans 2B to 32B plus a fast MoE, InternVL3.5 and Ai2’s Molmo2 arrived as serious open alternatives, and MiniCPM-V 4.6 redefined how much a sub-2B model can do. The tier picks below are rebuilt around them.
Drop a screenshot, a photo of a whiteboard, a scanned document, or a chart into a local AI model and get back text descriptions, extracted data, or answers to questions about what’s in the image. No cloud upload, no API costs, no privacy concerns.
Vision language models (VLMs) have improved dramatically. A model that fits on an 8GB GPU can now read documents more accurately than GPT-4o could two years ago. Here’s what works at every VRAM tier as of mid-2026.
The Benchmarks That Matter
- MMMU - college-level multi-discipline questions requiring image understanding. The gold standard for general visual reasoning
- DocVQA - document question answering. Can the model read a receipt, invoice, or form and answer questions about it?
- ChartQA - understanding charts and graphs. Crucial for data analysis workflows
- MathVista - mathematical reasoning from visual inputs (diagrams, equations, geometry)
- OCRBench - text extraction accuracy from images
- Pointing / grounding - can the model point to where an object is, not just say it exists? This is where Molmo2 specialises
What Changed Since Last Year
The Qwen2.5-VL / LLaVA / Llama-3.2-Vision generation that anchored older guides has been overtaken. Three families now lead the open-weight field:
- Qwen3-VL (released late 2025) - dense 2B / 4B / 8B / 32B plus a 30B-A3B mixture-of-experts, with 256K native context expandable toward 1M. The broadest and, on most tasks, the strongest open VLM family.
- InternVL3.5 (OpenGVLab) - 1B to 38B dense plus a large MoE flagship, adding a visual-resolution router the project credits for a roughly 4x inference speedup over InternVL3.
- Molmo2 (Ai2, fully open) - 4B and 8B models that became the open state of the art specifically for pointing and grounding and video tracking, surpassing the older 72B Molmo. Note the weights are Apache-2.0 but the training data is restricted to non-commercial use.
And at the small end, MiniCPM-V 4.6 (roughly 1 to 1.3B parameters) delivers far more capability per gigabyte than anything its size a year ago, running in about 4GB on GPU or ~2GB as a CPU GGUF build.
8GB VRAM {#8gb}
GPUs: RTX 4060, RTX 3060 8GB, RTX 3070
Candidates
| Model | Params | VRAM | Context | License | Notes |
|---|---|---|---|---|---|
| Qwen3-VL-8B | 8B | ~6 GB | 256K | Apache 2.0 | Flagship-family quality at consumer size |
| InternVL3.5-8B | 8.5B | 8-bit for consumer cards | long | Apache 2.0 | Strong reasoning, fast router |
| MiniCPM-V 4.6 | ~1-1.3B | ~4 GB (or ~2 GB CPU GGUF) | - | Apache 2.0 | Best capability-per-GB at the low end |
| Qwen3-VL-4B / -2B | 4B / 2B | ~3.3 / ~1.9 GB | 256K | Apache 2.0 | Fallback for tight budgets |
Winner: Qwen3-VL-8B
Qwen3-VL-8B is the default local vision model for most people. At roughly 6GB it fits an 8GB card with room for context, and it carries the accuracy of the wider Qwen3-VL family into consumer range. Point it at a screenshot and ask “what’s the error message?”, or photograph a whiteboard and get structured notes back. Document and OCR handling is genuinely good.
Reasoning alternative: InternVL3.5-8B. OpenGVLab reports a large reasoning gain and inference speedup over InternVL3; run it 8-bit to fit a consumer card. Worth a try if Qwen3-VL’s answers feel shallow on multi-step visual reasoning.
Ultralight pick: MiniCPM-V 4.6. At around 4GB on GPU (or ~2GB as a CPU GGUF) it is the smallest model here that still reads documents and charts usefully - ideal when you need vision running alongside a chat model, or on a machine with no discrete GPU.
For all use cases at this level, see our 8GB VRAM complete guide.
12GB VRAM {#12gb}
GPUs: RTX 3060 12GB, RTX 4070
Winner: Qwen3-VL-8B (with headroom)
Same model as the 8GB tier, now with roughly 6GB free for context. That headroom matters for vision: you can feed higher-resolution images and include more text alongside them, and inference runs faster. If you want a second opinion on hard reasoning shots, keep InternVL3.5-8B installed and switch between them.
Grounding specialist: Molmo2-8B. If your task is pointing at UI elements, locating objects, or tracking things across frames rather than general Q&A, Molmo2 is the open state of the art for that specific job. Ai2 does not publish a firm VRAM figure, so budget conservatively and test on your card.
For all use cases at this level, see our 12GB VRAM complete guide.
16GB VRAM {#16gb}
GPUs: RTX 4060 Ti 16GB, RTX 5060, Intel Arc A770, AMD RX 7800 XT
Here is an honest gap. The current open VLM families jump from roughly 6GB (the 8B-class models) straight to roughly 18-21GB (the 30B/32B-class models), so there is no model purpose-built for the 16GB bracket. The best use of 16GB is therefore not a bigger vision model but more room around the 8B one.
Winner: Qwen3-VL-8B at higher precision, plus a second model
Run Qwen3-VL-8B at a higher-quality quantization (Q6/Q8) for cleaner OCR of small text, and use the remaining VRAM to keep a small chat LLM or a grounding model (Molmo2-8B) resident at the same time. InternVL3.5-8B at full precision also fits comfortably here. You get sharper output and a two-model workflow rather than one starved big model.
For all use cases at this level, see our 16GB VRAM complete guide.
24GB VRAM {#24gb}
GPUs: RTX 3090, RTX 4090
Candidates
| Model | Params | VRAM | Notes |
|---|---|---|---|
| Qwen3-VL-32B | 32B dense | ~20-21 GB | Highest-quality single-GPU VLM at this tier |
| Qwen3-VL-30B-A3B (MoE) | 30B / 3B active | ~18-20 GB | Near-32B quality, faster inference |
| Qwen3-VL-8B | 8B | ~6 GB | Speed option, 100+ tok/s, leaves room for other models |
Winner: Qwen3-VL-32B
At 24GB you can run the big one. Qwen3-VL-32B at roughly 20-21GB pushes document understanding, chart reading, and visual reasoning to their strongest open-weight level while still fitting a single consumer card with headroom for context.
Speed vs quality: the Qwen3-VL-30B-A3B mixture-of-experts model lands in a similar VRAM footprint (~18-20GB) but only activates about 3B parameters per token, so it runs noticeably faster than the dense 32B at close to the same quality. For high-throughput OCR pipelines, prefer the MoE.
Dual model: run Qwen3-VL-8B (~6GB) for vision plus a chat model in the remaining space, both loaded at once - vision for images, the LLM for text conversations.
For all use cases at this level, see our 24GB VRAM complete guide.
32GB VRAM {#32gb}
GPUs: RTX 5090
Winner: Qwen3-VL-32B at higher quant, or a full multimodal stack
Same model as the 24GB tier, now with room to run it at a higher-quality quantization for fewer artifacts on small text and fine detail - the difference shows up in OCR of dense documents and complex diagrams.
Full vision pipeline: with 32GB you can run vision + chat + speech at once. Example: Qwen3-VL-8B (~6GB) for images + a 14B-class chat LLM (~11GB) + Parakeet/Kokoro for voice (~3GB) leaves margin to spare. Point a camera at something, ask about it, and hear the answer, all local.
For all use cases at this level, see our 32GB VRAM complete guide.
Cross-Tier Summary
| Tier | Best Pick | VRAM | Strength |
|---|---|---|---|
| 8GB | Qwen3-VL-8B | ~6 GB | Document OCR, charts, screenshots |
| 12GB | Qwen3-VL-8B + Molmo2 | ~6-8 GB | General + pointing/grounding |
| 16GB | Qwen3-VL-8B (Q6/Q8) + 2nd model | ~14 GB | Sharper output, two-model workflow |
| 24GB | Qwen3-VL-32B (or 30B-A3B MoE) | ~18-21 GB | Strongest single-GPU document analysis |
| 32GB | Qwen3-VL-32B (higher quant) | ~28 GB | Maximum quality + room for a full stack |
What Can Vision Models Actually Do?
The practical use cases that work well today:
- OCR and document reading - extract text from photos, scanned PDFs, receipts, invoices. The Qwen3-VL models are especially strong here
- Screenshot analysis - describe what’s on screen, identify UI elements, read error messages
- Chart and graph reading - extract data points, describe trends, answer questions about visualizations
- Pointing and grounding - locate a specific object or UI element by coordinates (Molmo2’s specialty)
- Whiteboard to text - photograph handwritten notes and get structured text output
- Math from images - photograph an equation or geometry problem and get it worked through
What they still struggle with:
- Spatial reasoning - “Is the red object to the left of the blue one?” is more reliable than a year ago but still trips up smaller models
- Counting - “How many people are in this photo?” often gets wrong answers on crowded scenes
- Fine-grained detail - very small text and subtle differences between similar objects
- Long video understanding - improving fast (Molmo2 tracks across frames), but still slow and memory-hungry at length
Quick Start
# Install Ollama and pull a vision model
ollama pull qwen3-vl:8b # 8GB tier
ollama pull qwen3-vl:32b # 24GB+ tier
# Analyze an image from the command line
ollama run qwen3-vl:8b "Describe this image" ./photo.jpg
# Or use the API
curl http://localhost:11434/api/chat -d '{
"model": "qwen3-vl:8b",
"messages": [{
"role": "user",
"content": "What text is in this image?",
"images": ["base64_encoded_image_here"]
}]
}'Qwen3-VL requires a recent Ollama build (0.12.7 or newer). For a web interface with image upload, pair Ollama with Open WebUI; we cover the setup in our Ollama + Open WebUI guide.