Nimbus8 Try Now
on-device · private · local-first

On-device intelligence.
Chat, code, and agents — all running locally, nothing leaves your device.

Download on the App Store
App StoreiPhone, iPad & Mac
iOS 26+Available now
100%on-device inference
Nimbus Studio

Eight modules. One native engine.

Explore the complete local-first workspace. Toggle between modules to see how they run on-device, and switch devices to preview native rendering on iPhone, iPad, and Mac.

Gale · Chat

A real chat app for open models.

Talk to instruct-tuned LLMs with the polish of a modern chat app. Streaming thoughts, structured tool calls, and local history. Zero cloud required.

  • Streaming repliestoken-by-token with collapsible thinking blocks
  • Private-firstno accounts, no remote servers, no telemetry
  • Curated catalogoptimized weights: Llama 3.2, Qwen3, Gemma 3, Phi-4
Deep dive into Gale
Cirrus · Code

From prompt to pull request.

Cirrus plugs straight into your GitHub repositories. Describe a task, inspect an inline diff card with hunk headers, and open a PR with a single tap.

  • Inline diffsstandard +/- code gutters, diff highlights, and counts
  • Repo-awareindexed local file contexts for precise generation
  • Model libraryQwen3-Coder (4B/7B), DeepSeek-Coder, CodeGemma
Deep dive into Cirrus
Ashe · Agent Runtime

Autonomous hands. Safe execution.

Ashe runs background tasks ("Hands") locally. Watch tool calls, checks, and models trigger in real time. Ashe always asks for permission before modifying files.

  • Activity streamlive audit log of every tick, read, and write
  • Confirm-firstsafety gate blocks destructive actions until approved
  • Local schedulerruns on ticks, wake triggers, or background cron
Deep dive into Ashe
Mist · Search

Offline knowledge + local retrieval.

Mist merges vector search and web results locally. Search public databases, extract text, and compile cited answers without leaking your query.

  • Offline packsdownload Wikipedia or stack exchange slices for offline use
  • Semantic embeddingspowered by BGE-small-en and EmbeddingGemma ONNX
  • Hybrid RRF fusionmerges BM25 keyword matching with dense vectors
Deep dive into Mist
Mirage · Image

Local diffusion on the Neural Engine.

Prompt-to-image running directly on Apple Silicon. Mirage leverages palettized diffusion models to render crisp images without consuming cloud tokens.

  • Neural Engine accelerationruns Core ML .mlmodelc packages at float16/6-bit
  • Curated modelsStable Diffusion 1.5 Pal, SDXL Turbo, and PixArt-α
  • Private sandboxgallery and generation history are strictly on-device
Deep dive into Mirage
Overture · Audio

Real-time local speech-to-text.

Dictate, transcribe recordings, and run voice-first workflows locally. Overture uses Whisper in GGML format for lightning-fast transcription.

  • Whisper pipelinesupports Whisper Tiny, Base, Small, and Medium
  • Word highlightsextracts timestamps from attention weights for overlays
  • Lightweight TTSsynthesis via Kokoro-82M ONNX at 150 words/sec
Deep dive into Overture
Stratus · Vision

Visual parsing and local OCR.

Point your camera, capture, and extract structured data. Stratus routes images through multimodal chat models and fast OCR kernels.

  • Capture-to-askquery details about photos, text pages, or physical objects
  • Local extractionhighly accurate text recognition with zero latency
  • Multimodal routingtargets vision-enabled models via capability manifest
Deep dive into Stratus
Zephyr · Translate

Local translation with chat models.

Translate documents, notes, or messages side-by-side. Zephyr utilizes chat LLMs with translation templates for context-aware, idiomatic translations.

  • Context-awaregoes beyond simple word matching to preserve grammar
  • Wide language supportsupports dozens of language pairs verified by test suites
  • Model flexibilityruns via Llama 3.2 3B, Qwen3 4B, or Gemma 3 4B
Deep dive into Zephyr
Nimbus8
Create
Gale
Mirage
Stratus
Mist
Zephyr
Sound
Overture
Pro
Cirrus
Ashe
Spire
Library
Haze
History
Models
Projects
9:41
Nimbus8
Good afternoon, Cameron
Where should we start?
Who won the men's 100m in the Paris 2024 Olympics?
Web Search
searching Wikipedia
Ask Nimbus8
9:41
Cirrus
anthropics/claude-cookbook
base: main
›
Find the rate-limit bug in requests.py
Read requests.py
128 lines scanned
Found it — the exponential backoff resets on every retry instead of compounding. Here's a patch:
requests.py +4 −2
@@ -42,6 +42,8 @@ def fetch_with_retry(url):
42for attempt in range(max_retries):
43 try:
44 return http.get(url)
45− except RateLimit:
46− time.sleep(1)
45+ except RateLimit:
46+ delay = base * (2 ** attempt)
47+ time.sleep(delay + jitter())
48+ continue
Message Cirrus…
9:41
Ashe
Activity
Live tool calls, model calls, and hand ticks.
Invoked skill · inbox.triage
3 emails classified
Ran hand · morning-brief
2.4s · model: Gemma-3-4B
Thinking
drafting summary…
Needs confirmation
write to ~/Nimbus8/brief.md
Hands
3 enabled · next tick in 6m
morning-brief
every 1h · last: 2m ago · 0 errors
inbox.triage
every 15m · last: 14m ago · 0 errors
9:41

Mist

capability manifest routing
YOUR NOTES 2
capability_manifest.md
0.91

The capability manifest lists each model's vision, PDF, OCR, and tool-use flags. Routing picks the model that both handles the input type and fits your device's memory budget — quantization is preferred when the tier budget is tight.

ModelRouter.swift
0.84

ModelRouter.classify() takes a repoId, filename, tags, libraryName, and pipelineTag and returns a Classification with .module and .confidence.

WEB 1
Running open models on iPhone — Nimbus8 blog
trynimbus.org

A quick tour of MLX vs GGUF on Apple Silicon, how we rank quants, and why a 3B-Q4 often beats a 7B in day-to-day chat.

42 sources · 768-dim 38 ms
9:41
Mirage
PROMPT
A lone cirrus cloud over a warm desert at golden hour, cinematic, 35mm, soft grain.
Step 14 of 20
~6s
SD 1.5 Pal 512²
9:41
Overture
Recording · 00:42 WhisperKit · base
We're shipping the agent runtime this quarter — the capability manifest is done and routing's already landed. What's blocking us is the tool-use flags for the smaller models; I'd rather ship with three tools than ship late with six.
9:41
Stratus
Recognized text · 98%
CAUTION — PRIVATE ROAD NO THROUGH TRAFFIC Speed limit 15
A red-and-white private-road sign on a wooden post, cropped near a curved driveway. Text reads "Caution — Private Road, No Through Traffic, Speed limit 15".
Ask about this photo…
9:41
Zephyr
ENGLISH
Run open-source language models locally on your Apple devices.
SPANISH
Template: Instruct LLM
Hardware Compatibility

Find the model that fits your device.

Because Nimbus8 runs entirely on-device, memory is the ultimate budget. Slide below to select your device RAM and see which models run optimally. We account for both raw weights and active KV cache overhead to prevent Out-Of-Memory (OOM) crashes.

6 GB RAM
iPhone 15, iPad 10th Gen 8 GB RAM
iPhone 15 Pro, iPhone 16/16 Pro, M1 iPad 12 GB RAM
iPhone 17, iPhone 17 Pro 16 GB+ RAM
M-series Mac, M4 iPad Pro

Llama 4 1B

Highly optimized ultra-lightweight 2026 chat model.

QuantMLX 4-bit
Size0.6 GB
Speed~65 tok/s

Gemma 3 1B

Google's mid-2026 lightweight model with deep reasoning.

QuantMLX 4-bit
Size0.7 GB
Speed~58 tok/s

Qwen3 1.7B

Excellent reasoning and multilingual capabilities for compact devices.

QuantMLX 4-bit
Size1.1 GB
Speed~48 tok/s

Llama 4 3B

Premium balanced language assistant, fits well with active KV cache overhead.

QuantMLX 4-bit
Size1.8 GB
Speed~38 tok/s

Gemma 3 4B

State-of-the-art mid-2026 chat model with visual capability.

QuantMLX 4-bit
Size2.6 GB
Speed~32 tok/s

Qwen3 4B

Robust general assistant, runs comfortably on 8 GB devices.

QuantMLX 4-bit
Size2.4 GB
Speed~35 tok/s

Qwen3-Coder 7B

Top-tier coding model. Demands 12 GB+ RAM to accommodate 1.8 GB KV cache.

QuantMLX 4-bit
Size4.4 GB
Speed~25 tok/s

DeepSeek-R1 Qwen 7B

Distilled reasoning model with offline chain-of-thought capabilities.

QuantQ4_K_M GGUF
Size4.3 GB
Speed~24 tok/s

Llama 4 8B

Flagship local model. Requires 16 GB+ RAM for full 8k context window.

QuantQ4_K_M GGUF
Size4.9 GB
Speed~20 tok/s

Now live on the App Store.

Download Nimbus8 today for iPhone, iPad, and Mac. Gale is free forever, and Nimbus8 Pro unlocks the full local studio with a one-time Apple In-App Purchase.

Download on the App Store View Pro pricing

Available now for iOS 26+, iPadOS 26+, and macOS 26+ on Apple silicon Macs.