Where to find us
The Nimbus8 community lives in three places, each suited to a different kind of conversation. GitHub Discussions (github.com/trynimbus) is the home for bug reports, feature requests, and technical questions about the runtime. If something crashed, didn't load, or behaved unexpectedly, that's where to start. Issues are triaged weekly and tagged by module.
For model recommendations and benchmarking conversations, head to the Hugging Face community space (huggingface.co/trynimbus). This is where people share what models they've tested, how they performed on specific devices, and which quantization variants worked best. It's a good place to browse before downloading something new.
Discord (discord.gg/trynimbus) is for real-time chat. Ask a quick question, share a screenshot of something cool you generated with Mirage, or just hang out with other people who care about running AI locally. The server has channels organized by module, plus a general channel for off-topic conversation. We're a small community, so don't be surprised if a core contributor answers your question directly.
Model recommendations
The curated catalog covers the models we've tested and verified, but the community often finds great picks that haven't made it into the official list yet. Community model recommendations are shared on the Hugging Face space with a simple format: model name, quantization, device tested, module used, and a brief note on quality and speed.
If you've found a model that works well for a specific task — say, a fine-tuned Llama variant that's great at summarization, or a Whisper derivative that handles accented English better than the base models — share it. The best community picks get promoted to the curated catalog in the next release after we verify them on our reference devices.
When sharing benchmarks, include your device model (e.g., iPhone 15 Pro), the model file you downloaded (repo + filename), the module you tested in, and the tok/s or generation time you observed. Screenshots of the model info screen in Nimbus8 are helpful too — it shows memory usage, context length, and backend in one place.
Device reports
One of the most valuable things you can contribute is a device report. Every iPhone and iPad has different thermal characteristics, memory pressure behavior, and Neural Engine throughput. A model that runs at 30 tok/s on an iPhone 15 Pro might only manage 18 tok/s on an iPhone 14 Pro due to the A16 vs A17 Pro difference — and that's useful information for everyone.
The format is simple: device model, iOS version, model name + quantization, module, tok/s (or generation time for images), and any notes about thermal throttling or memory warnings. Post these on GitHub Discussions under the "Device Reports" category. We aggregate them into a compatibility matrix that informs the catalog's device-fit gating.
If you're on an older device — iPhone 12, iPhone 13 mini, iPad Air 4 — your reports are especially valuable. We test primarily on iPhone 15 Pro and iPad Pro M2, so edge-of-support devices are where community data fills the biggest gaps. Even a report that says "this model didn't load" or "the app got killed by the OS after 30 seconds" helps us set better memory thresholds.
Contributing
The Nimbus8 iOS app is closed-source, but the Rust core library and this landing site are open. If you want to contribute code, the Rust core is where most of the interesting work happens — the inference client, diff engine, download manager, embedding pipeline, and safety scanner all live there. The repo is at github.com/trynimbus/nimbus-core.
Good first issues are tagged in the repo. They tend to be things like adding a new model format parser, improving error messages, or writing tests for edge cases in the diff engine. If you want to tackle something bigger, open a discussion first so we can align on the approach before you invest time.
Bug reports are contributions too. A good bug report includes: what you did, what you expected, what happened instead, your device model and iOS version, and the model you were using. Crash logs from Settings → Privacy → Analytics are incredibly helpful — they include the stack trace that tells us exactly where things went wrong. If you can reproduce the issue reliably, that's gold.
Community guidelines
Be kind. Everyone here is figuring things out — from people running their first local model to contributors debugging FFI edge cases. Ask questions without hesitation, answer questions without condescension, and assume good intent. We're building something together, and that works best when everyone feels welcome.
Stay on topic. The community spaces are for Nimbus8, local AI, open models, and related technical discussion. Political debates, cryptocurrency promotion, and unrelated self-promotion don't belong here. We keep the signal-to-noise ratio high so that when you visit, you find useful information quickly.
Respect model licenses. Many open-weight models come with specific license terms — Llama has the Meta Community License, Gemma has Google's terms, Mistral has Apache 2.0. Don't share model files directly; link to the Hugging Face repo so people download through the proper channel. Don't share techniques for bypassing model safety filters or prompt injection methods. We take this seriously and will remove content that violates these norms.
FAQ
Why no Android? Nimbus8 is built on Apple's MLX framework, Core ML, and the Neural Engine — none of which exist on Android. A port would mean rewriting the entire inference stack, not just the UI. We're a small team and chose to go deep on one platform rather than shallow on two. If Android support becomes feasible, we'll announce it here first.
Why no cloud option? The whole point of Nimbus8 is that nothing leaves your device. Adding a cloud inference option would compromise that promise and introduce account management, API keys, billing, and data handling policies that we don't want to build or maintain. If you want cloud inference, there are excellent apps for that — we're building something different.
Can I use my own server? Yes. Nimbus8 supports custom OpenAI-compatible endpoints for any module. Point it at a local Ollama instance, a vLLM server on your home network, or any API that speaks the OpenAI chat completions format. HTTPS is enforced for all endpoints except localhost/loopback addresses. Configure this in Settings → Custom Endpoints.