| Scope and supported models | A general runner for open models. The README names Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models, and points to the ollama.com library for the full list, which includes qwen3.8-flash-next and qwen3.6. | An engine for one model family. Its model guide covers Qwen3.8-Flash-Next, in ISTA-DASLab, Unsloth and other packs, and the smaller Qwen3.6-35B-A3B. Strata no longer runs DeepSeek V4.1-Flash: version 0.2.13 was the last one that set it up and started it. |
|---|
| How models are obtained | ollama run and ollama pull fetch models from the ollama.com library. A Modelfile builds customized models, and the import guide covers Safetensors weights and GGUF files, which Ollama does not quantize during import. Models you create can be pushed to ollama.com. | You choose a model in the desktop app or in strata launch, and the files download from Hugging Face, with ModelScope as a second source. strata models import shows what files you already have would become. Since 0.2.15 Strata loads models only on its own engine. |
|---|
| Hardware and memory | NVIDIA GPUs with compute capability 5.0 and newer, AMD through ROCm v7, Apple GPUs through Metal, and more GPUs through Vulkan on Windows and Linux. A model that does not fit on one GPU is spread across all available GPUs, and Ollama can also run on the CPU from system memory. | The graphics card holds as many experts as its VRAM allows, RAM holds all of them and the SSD holds a lookup table. With 48 GB of RAM, Q2_0 and IQ2_XS fit; with 64 GB, every ISTA-DASLab size fits. A supported graphics card is required: an Intel UHD or Iris chip alone is not enough. |
|---|
| Platforms | Windows 10 22H2 or newer, macOS Sonoma (14) or newer, where x86 Macs run on the CPU only, and Linux on x86-64 and ARM64, with an extra ROCm package for AMD GPUs. The official ollama/ollama Docker image is on Docker Hub. | Windows 10 and 11 and Linux x86-64, macOS 15 or newer on Apple Silicon, and Linux ARM64 on NVIDIA DGX Spark through the Terminal only. Intel Arc runs on Linux only. There is one installer per system, and the app downloads the engine for your graphics card on first start. |
|---|
| APIs and tool integrations | A native REST API on port 11434 plus subsets of the OpenAI API and the Anthropic Messages API, and official libraries such as the Python package. ollama launch supports Claude Code, Codex, Copilot CLI, DeepSeek Harness, Droid and OpenCode, and the README lists many community integrations such as Open WebUI. | The server on http://127.0.0.1:8080 speaks OpenAI Chat Completions, OpenAI Responses and Anthropic Messages, each with streaming and tools, and dashboards written for vLLM can read it. strata connect covers OpenCode, Codex CLI, Claude Code, Hermes Agent, OpenClaw and others, and can send one prompt and one tool call the way each app does. |
|---|
| Privacy and cloud | Ollama binds 127.0.0.1 port 11434 by default and says it does not see your prompts or data when you run locally. Cloud models run on Ollama's cloud service and require an account; OLLAMA_NO_CLOUD=1 turns on a local-only mode without cloud models and web search. | There is no telemetry and no account. The server listens on your PC only, another device needs an API key, and once a model is downloaded Strata runs without internet. In the app you can add outside services such as OpenAI, Anthropic or OpenRouter. |
|---|
| Licence and pricing | MIT License, copyright Ollama. The free plan includes running models locally and charges per token for cloud models, and paid plans start at $20 a month or $200 a year. Check each model's licence before commercial use. | MIT License, copyright Niko1221 and the Strata contributors. A few bundled parts carry their own licences, the models are not part of the repository, and each model's own licence applies to its files. Free to run, with no account needed. |
|---|
| Maturity and releases | The longer track record: the Ollama blog has posts from 2023, including the announcement of the official Docker image. Version numbers have reached v0.40.16, released on 9 November 2026, and v0.40.0 made Apple Silicon run MLX-supported model architectures on MLX automatically. | Young and fast moving. The first release, v0.1.0 on 24 September 2026, was an engine for Windows with RTX 30, 40 and 50 cards. Version 0.2.15 followed on 16 November, a day after 0.2.14, and its notes list known problems such as Radeon issues on Linux. |
|---|