Models
This website cannot detect your RAM, GPU or installed models. Nothing below is marked Compatible. The installed app inspects hardware after launch and only then assigns READY, POSSIBLE or INCOMPATIBLE.
On this website
AI setup required
Local AI is not claimed ready in the browser. If a phone cannot run a local runtime, the app should say so and offer cloud only if you configure a provider.
ARM64, typically 6–12 GB RAM. iOS often hides RAM from apps — that value is marked UNKNOWN, never guessed.
Reasoning
Llama 3.2 1B, Gemma 2 2B or Phi-3 Mini
Small enough for on-device inference when a mobile runtime is present. Desktop 7B+ models are not assumed to run here.
OCR
On-device OCR / camera capture
Read a page photo before any reasoning model runs.
Embeddings
all-minilm if RAM allows
Tiny retrieval only. Skip if the device is tight.
A phone is not a workstation. The app must not recommend Mixtral, 14B or 32B models on a handset.
ARM64, often more RAM and cooling headroom than a phone. iPad RAM may still be unavailable to the app.
Reasoning
Llama 3.2 3B or Qwen2.5 3B
Slightly larger than phone-class when memory is actually reported and a runtime is present.
Vision
Moondream if scans include diagrams
Optional. Printed text still goes through OCR first.
Embeddings
nomic-embed-text
Useful for longer packs on a tablet.
Tablet recommendations stay in the mobile-suitable catalog. Ollama desktop pulls are not treated as iPad installs.
Catalog size class only. Not installed. Not Compatible until the app verifies a runtime on that device.
Llama 3.2 1B
1B · ~1.3 GB · ~2 GB + OS
Smallest Llama option for short questions on phones and low-RAM PCs.
Not verified on this phone
Llama 3.2 3B
3B · ~2 GB · ~4 GB + OS
Common first local model for cited Q&A on tablets and modest PCs.
Not verified on this phone
Gemma 2 2B
2B · ~1.6 GB · ~3 GB + OS
Very small instruction model for phones and tight RAM.
Not verified on this phone
Phi-3 Mini
3.8B · ~2.3 GB · ~4 GB + OS
Compact reasoning for phones and tablets; needs chunking on long files.
Not verified on this phone
Qwen2.5 3B
3B · ~2 GB · ~4 GB + OS
Small Qwen with a longer window than most tiny models. Suitable for tablets when RAM is reported.
Not verified on this phone
Moondream
1.8B · ~1.7 GB · ~4 GB + OS
Small vision model for stamps or diagrams. Not OCR and not a contract reasoner.
Not verified on this phone
nomic-embed-text
137M · ~0.3 GB · ~1 GB
Local embeddings for retrieving passages. Does not answer questions.
Not verified on this phone
all-minilm
23M · ~0.05 GB · ~1 GB
Smallest embedding option for phones and low-RAM devices.
Not verified on this phone
Not assumed on Android or iPhone. Do not treat these as mobile-compatible.
Gemma 2 9B
9B · Desktop / workstation only
Qwen2.5 7B
7B · Desktop / workstation only
Mistral 7B
7B · Desktop / workstation only
Llama 3.1 8B
8B · Desktop / workstation only
DeepSeek R1 8B
8B · Desktop / workstation only
Mistral Nemo
12B · Desktop / workstation only
Qwen2.5 14B
14B · Desktop / workstation only
Phi-4
14B · Desktop / workstation only
DeepSeek R1 14B
14B · Desktop / workstation only
Qwen2.5 32B
32B · Desktop / workstation only
Mixtral 8x7B
8x7B · Desktop / workstation only
LLaVA 7B
7B · Desktop / workstation only
Llama 3.2 Vision
11B · Desktop / workstation only
mxbai-embed-large
335M · Desktop / workstation only
BGE-M3
568M · Desktop / workstation only
Adaptive local AI
The app recommends a single model or a specialised stack after it scans this device — Windows, Mac, Linux, Android or iPhone/iPad. Desktop Ollama pulls are not treated as phone installs. Sizes below are typical Q4 estimates. Nothing is installed until you approve it.
ContractAnalysis.ai — Your AI Environment
DemonstrationFirst-run setup
ContractAnalysis.ai has analysed this device and identified the best AI options available. Demonstration figures below are labelled examples, not a live scan of your phone or computer.
CPU
Example 8-core processor
System RAM
Example 32 GB
GPU
Example dedicated GPU
AI runtime
Ollama — Connected
Recommended
A mid-size local reasoning model that fits reported memory
Demonstration interface. The installed app scans the actual computer and never lists a model as installed unless a local runtime reports it.
These are planning guides. The installed app uses the hardware and runtimes it actually detects. A smaller model that runs is preferred to a larger one that will run out of memory.
Basic PC
About 8 GB RAM, CPU-only or modest integrated graphics
Fits limited memory and still answers short, cited questions.
Local scan text without a vision model.
Tiny retrieval only. Skip if RAM is tight.
A single small reasoning model is preferred. Do not install 14B+ models on this class of machine.
Capable PC
About 16 GB RAM, optional dedicated GPU
Better document Q&A while remaining realistic on 16 GB.
Passage retrieval for larger files.
Default for scans.
One capable reasoning model is usually enough. Add embeddings if you work with long packs.
High-performance PC
About 32 GB RAM and a dedicated GPU with reported VRAM
Stronger reasoning over larger excerpt sets.
Optional. OCR still handles printed text.
Better retrieval across a document set.
A specialised stack is optional. The app should not force extra models if one 14B model already fits.
Workstation
48 GB+ RAM and substantial dedicated VRAM
Highest quality in this catalog when memory allows.
Mixed text and image pages.
Indexing large multi-file packs.
Do not download enormous models just because the machine can. Recommend only what you will use.
Android phone or iPhone
ARM64, typically 6–12 GB RAM. iOS often hides RAM from apps — that value is marked UNKNOWN, never guessed.
Small enough for on-device inference when a mobile runtime is present. Desktop 7B+ models are not assumed to run here.
Read a page photo before any reasoning model runs.
Tiny retrieval only. Skip if the device is tight.
A phone is not a workstation. The app must not recommend Mixtral, 14B or 32B models on a handset.
Android tablet or iPad
ARM64, often more RAM and cooling headroom than a phone. iPad RAM may still be unavailable to the app.
Slightly larger than phone-class when memory is actually reported and a runtime is present.
Optional. Printed text still goes through OCR first.
Useful for longer packs on a tablet.
Tablet recommendations stay in the mobile-suitable catalog. Ollama desktop pulls are not treated as iPad installs.
Android and iPhone/iPad are included in the same licence. Download the package for your device after purchase. Local inference on a phone uses only mobile-suitable models — typically 1B–3B class — and only after a compatible runtime is detected. A 14B desktop model is marked incompatible on a handset. If iOS does not expose RAM to the app, that field is UNKNOWN and the recommendation stays conservative.
Packages: Windows 10 or 11 (64-bit) · Android 10 or later (64-bit) · iPhone and iPad (iOS / iPadOS 16 or later) · macOS 12 or later (Apple Silicon or Intel) · Linux (x86_64 or ARM64).
Status is a detection policy, not a claim that every device already has a model running.
| Capability | All supported platforms |
|---|---|
| Local AI | Detected on this device — never assumed |
| Model discovery | Catalog plus what a reachable runtime actually lists |
| Local models | Runtime and hardware dependent |
| Vision | Only if a vision-capable model is present and tested |
| Embeddings | Yes where a compatible embedding model is configured |
| Cloud AI | Optional, explicit, never the silent default |
Catalog entries are downloadable options when a supported runtime is running. They are not preinstalled. A model is not available on a phone just because it exists in this table.
| Model | Role | Typical device | Mobile | Download | Est. RAM | Abilities |
|---|---|---|---|---|---|---|
| Llama 3.2 1B llama3.2:1b | reasoning | Phone | Suitable when a mobile runtime is present | ~1.3 GB | ~2 GB + OS | Smallest Llama option for short questions on phones and low-RAM PCs. Strengths: Low memory · Fast on CPU Limits: Weak on long contracts |
| Llama 3.2 3B llama3.2:3b | reasoning | Tablet | Suitable when a mobile runtime is present | ~2 GB | ~4 GB + OS | Common first local model for cited Q&A on tablets and modest PCs. Strengths: Fits many 8 GB machines · Clear short answers Limits: Limited nuance |
| Gemma 2 2B gemma2:2b | reasoning | Phone | Suitable when a mobile runtime is present | ~1.6 GB | ~3 GB + OS | Very small instruction model for phones and tight RAM. Strengths: Small download Limits: Shallow multi-document reasoning |
| Phi-3 Mini phi3:mini | reasoning | Phone | Suitable when a mobile runtime is present | ~2.3 GB | ~4 GB + OS | Compact reasoning for phones and tablets; needs chunking on long files. Strengths: Better logic than 1–2B class Limits: Short context |
| Qwen2.5 3B qwen2.5:3b | reasoning | Tablet | Suitable when a mobile runtime is present | ~2 GB | ~4 GB + OS | Small Qwen with a longer window than most tiny models. Suitable for tablets when RAM is reported. Strengths: Longer context than Phi-3 Mini Limits: Still a small model |
| Gemma 2 9B gemma2:9b | reasoning | Capable PC | Desktop / workstation only | ~5.4 GB | ~10 GB + OS | Mid-size general model for clearer multi-clause answers. Strengths: Better instruction quality than 2–3B Limits: Needs a 16 GB-class PC |
| Qwen2.5 7B qwen2.5:7b | reasoning | Capable PC | Desktop / workstation only | ~4.7 GB | ~8 GB + OS | Balanced local reasoning model for document excerpts. Strengths: Long context · Structured answers Limits: Slower on CPU-only |
| Mistral 7B mistral:7b | reasoning | Capable PC | Desktop / workstation only | ~4.1 GB | ~8 GB + OS | General 7B instruction model. Strengths: Clear short prose Limits: Shorter context than Qwen2.5 7B |
| Llama 3.1 8B llama3.1:8b | reasoning | Capable PC | Desktop / workstation only | ~4.9 GB | ~10 GB + OS | Strong general model. Do not load a whole 200-page PDF at once. Strengths: Broad instruction quality Limits: Long context is not free RAM |
| DeepSeek R1 8B deepseek-r1:8b | reasoning | Capable PC | Desktop / workstation only | ~4.9 GB | ~10 GB + OS | More deliberate reasoning on supplied excerpts. Strengths: Useful for “what differs” questions Limits: Can be slower |
| Mistral Nemo mistral-nemo | reasoning | High-performance | Desktop / workstation only | ~7.1 GB | ~12 GB + OS | Larger Mistral model when RAM allows. Strengths: Richer long-form answers Limits: Heavy for 16 GB machines |
| Qwen2.5 14B qwen2.5:14b | reasoning | High-performance | Desktop / workstation only | ~9 GB | ~16 GB + OS | Stronger reasoning for large excerpt sets. Strengths: Better clause discrimination Limits: Poor experience under ~32 GB |
| Phi-4 phi4 | reasoning | High-performance | Desktop / workstation only | ~9.1 GB | ~16 GB + OS | Higher-quality reasoning model. Still needs grounded excerpts. Strengths: Structured answers Limits: Heavy for laptops |
| DeepSeek R1 14B deepseek-r1:14b | reasoning | High-performance | Desktop / workstation only | ~9 GB | ~16 GB + OS | Heavier reasoning for comparison of cited passages. Strengths: Useful for contradictions Limits: Slow without GPU |
| Qwen2.5 32B qwen2.5:32b | reasoning | Workstation | Desktop / workstation only | ~20 GB | ~28 GB + OS | Large local model. Only when the machine can hold it. Strengths: Highest quality in this catalog when it fits Limits: Very large download |
| Mixtral 8x7B mixtral | reasoning | Workstation | Desktop / workstation only | ~26 GB | ~32 GB + OS | Mixture-of-experts. Often incompatible on ordinary PCs. Strengths: Strong general quality when it runs Limits: Huge RAM and download |
| Moondream moondream | vision | Tablet | Suitable when a mobile runtime is present | ~1.7 GB | ~4 GB + OS | Small vision model for stamps or diagrams. Not OCR and not a contract reasoner. Strengths: Tiny vision option Limits: Not a replacement for Tesseract |
| LLaVA 7B llava:7b | vision | Capable PC | Desktop / workstation only | ~4.7 GB | ~10 GB + OS | Vision-language model for page images. Strengths: Can describe visual page content Limits: Weak as a text-only contract model |
| Llama 3.2 Vision llama3.2-vision | vision | High-performance | Desktop / workstation only | ~7.9 GB | ~14 GB + OS | Vision-capable Llama for mixed scans. Strengths: Vision plus some reasoning Limits: Prefer selectable text when it exists |
| nomic-embed-text nomic-embed-text | embedding | Phone | Suitable when a mobile runtime is present | ~0.3 GB | ~1 GB | Local embeddings for retrieving passages. Does not answer questions. Strengths: Tiny · Good retrieval helper Limits: Not a reasoning model |
| mxbai-embed-large mxbai-embed-large | embedding | Capable PC | Desktop / workstation only | ~0.7 GB | ~1 GB | Larger embedding model for retrieval. Strengths: Stronger retrieval than MiniLM-class Limits: Cannot analyse contracts alone |
| BGE-M3 bge-m3 | embedding | Capable PC | Desktop / workstation only | ~1.2 GB | ~2 GB | Multilingual embedding model. Strengths: Multilingual retrieval Limits: Not a Q&A model |
| all-minilm all-minilm | embedding | Phone | Suitable when a mobile runtime is present | ~0.05 GB | ~1 GB | Smallest embedding option for phones and low-RAM devices. Strengths: Tiny download Limits: Weaker retrieval |
On first launch the application scans reported platform, CPU architecture, RAM (or marks RAM UNKNOWN), GPU, NPU, storage and any reachable local runtime. It then marks each catalog model READY, POSSIBLE, NOT RECOMMENDED or INCOMPATIBLE for that device. You can use the recommendation, choose another compatible model, or continue with extraction only. See local AI setup.
Sizes and performance notes are estimates from published model size and typical hardware, not measured benchmarks on your machine.
