Run AI on your own machine
adris.tech can use a model that lives on your computer instead of the cloud. Nothing leaves your laptop, there is no usage meter to watch, and it keeps working offline.
A local model can do everything the hosted AI does. Web lookups, Google Maps, LinkedIn scanning and file handling are performed by the app, not by the cloud — a local model drives exactly the same tools. What changes is size: writing and summaries run happily on a small model, while lookups, multi‑step jobs and research need a larger one. adris.tech checks your memory and free disk, then tells you which model fits.
51 models · 12 free on every plan · 33 more on paid plans · 6 frontier models via Mesh
Free on every plan
All 12 models below are free to download and run, including on the Free plan. No usage meter — they run on your hardware.
| Model | Download | Memory needed | Best for |
|---|---|---|---|
| Llama 3.1 8B Free | 4.9 GB | 6 GB | Meta flagship small model |
| Qwen 2.5 Coder 7B Free | 4.3 GB | 6 GB | Top coder at 7B — beats models 2× its size |
| Mistral 7B Free | 4.4 GB | 6 GB | Best quality-to-size at 7B |
| Gemma 2 9B Free | 5.7 GB | 8 GB | Previous Gemma 2 generation |
| DeepSeek R1 7B Free | 4.5 GB | 6 GB | Strong chain-of-thought reasoning at 7B |
| Sarvam-1 7B Free | 4.6 GB | 6 GB | India-first model |
| DeepSeek Coder 6.7B Free | 3.9 GB | 6 GB | Code-specialised |
| Gemma 3 4B Free | 3.0 GB | 4 GB | Google compact model |
| Phi-4 Mini Free | 2.4 GB | 4 GB | Runs on 4 GB RAM, no GPU needed |
| Qwen 2.5 3B Free | 2.0 GB | 4 GB | Solid coding model for CPU-only devices |
| Llama 3.2 3B Free | 2.0 GB | 4 GB | Meta small instruction model |
| Llama 3.2 1B Free | 0.8 GB | 2 GB | Meta ultra-tiny |
Paid plans — all 33 models
Every model a single machine can run, largest first. The bigger ones handle tool-driven work — web lookups, multi-step runs and research — entirely on your own computer. Paid plans can also import any model file you already have.
| Model | Download | Memory needed | Best for |
|---|---|---|---|
| Qwen 2.5 72B Paid | 44 GB | 48 GB | Outperforms Llama 3 |
| DeepSeek R1 70B Paid | 43 GB | 48 GB | Near o1 quality reasoning on maths, science, and complex analysis |
| Llama 3.3 70B Paid | 43 GB | 48 GB | Meta flagship |
| Mixtral 8×7B Paid | 26 GB | 32 GB | 8-expert MoE (7B each) |
| Falcon 40B Paid | 24 GB | 32 GB | TII open-source 40B |
| Code Llama 34B Paid | 20 GB | 24 GB | Meta code specialised 34B |
| Command R 35B Paid | 20 GB | 24 GB | Cohere flagship with RAG and tool-use capabilities |
| Yi 1.5 34B Paid | 20 GB | 24 GB | Strong multilingual 34B from 01 |
| Qwen 2.5 Coder 32B Paid | 19 GB | 24 GB | Best open code model at 32B — beats GPT-4 on several coding benchmarks |
| DeepSeek R1 32B Paid | 19 GB | 24 GB | R1 reasoning at 32B |
| QwQ 32B Paid | 19 GB | 24 GB | Alibaba reasoning model |
| Qwen 2.5 32B Paid | 19 GB | 24 GB | Top-tier coding and reasoning, near 70B quality at half the size |
| Aya Expanse 32B Paid | 19 GB | 24 GB | Multilingual model covering 23 languages including Hindi, Arabic, Turkish |
| Gemma 3 27B Paid | 16 GB | 20 GB | Google powerful Gemma 3 |
| Mistral Small 24B Paid | 14 GB | 16 GB | Production-grade 24B with 128K context |
| Phi-4 22B Paid | 14 GB | 16 GB | Near-GPT-4 reasoning at 22B |
| Codestral 22B Paid | 12 GB | 16 GB | Mistral dedicated code model |
| InternLM 2.5 20B Paid | 12 GB | 16 GB | 1 million token context window |
| DeepSeek Coder V2 16B Paid | 9.4 GB | 12 GB | MoE architecture — 16B active but near 30B+ quality on code |
| StarCoder2 15B Paid | 9.0 GB | 12 GB | 600+ programming languages |
| Phi-4 14B Paid | 8.5 GB | 12 GB | Competes with 70B models on reasoning |
| Qwen 2.5 14B Paid | 8.5 GB | 12 GB | Strong all-round 14B with 128K context |
| DeepSeek R1 14B Paid | 8.5 GB | 12 GB | Larger R1 distillation |
| Gemma 3 12B Paid | 7.5 GB | 10 GB | Google mid-size Gemma 3 |
| Mistral Nemo 12B Paid | 7.2 GB | 10 GB | Mistral + NVIDIA collaboration |
| Llama 3.2 11B Vision Paid | 7.0 GB | 10 GB | Meta multimodal model — understands images alongside text |
| SOLAR 10.7B Paid | 6.5 GB | 8 GB | Outperforms Llama 2 70B despite being 10B, using depth upscaling |
| Gemma 3 9B Paid | 5.7 GB | 8 GB | Google compact mid-size |
| Phi-3.5 Mini Paid | 2.4 GB | 4 GB | 128K context at 3 |
| Gemma 2 2B Paid | 1.5 GB | 2 GB | Tiny but capable |
| SmolLM2 1.7B Paid | 1.1 GB | 2 GB | Built for on-device use |
| Qwen 2.5 1.5B Paid | 1.0 GB | 2 GB | Runs on 2 GB RAM |
| TinyLlama 1.1B Paid | 0.7 GB | 2 GB | Smallest model in the hub |
| Your own .gguf file Paid | any | — | Import any model file you already have |
Mesh — 6 frontier models
Too large for one computer. Mesh splits them across several machines on your network.
| Model | Download | Memory needed | Best for |
|---|---|---|---|
| DeepSeek R1 671B Mesh | 404 GB | 448 GB | Full DeepSeek R1 — o1 level reasoning |
| DeepSeek V3 Mesh | 404 GB | 448 GB | DeepSeek 671B MoE frontier |
| Llama 3.1 405B Mesh | 244 GB | 256 GB | Meta frontier model |
| DeepSeek V2 236B Mesh | 142 GB | 160 GB | Previous DeepSeek flagship MoE |
| Mixtral 8×22B Mesh | 87 GB | 96 GB | 8-expert MoE at 22B each |
| Mistral Large 123B Mesh | 74 GB | 80 GB | Mistral flagship 123B with 128K context |
Sizes are the quantised downloads adris.tech uses. The app never suggests a model your machine cannot comfortably run — it checks free disk and memory first.