Ishi Meets DeepSeek R1

DeepSeek's R1 is making waves as one of the most capable open reasoning models available. Today, we're excited to share how you can run R1 locally with Ishi — no cloud dependency, full privacy, and complete Glass Box transparency.

Why DeepSeek R1?

DeepSeek R1 brings reasoning capabilities previously only available in proprietary models:

• Chain-of-thought reasoning — Watch the model think step-by-step
• Open weights — Run locally via Ollama or other providers
• Competitive performance — Matches GPT-4o on many benchmarks
• Zero data leakage — Your files never leave your machine

Setting Up R1 with Ishi

Option 1: Ollama (Recommended for Local)

# Install Ollama if you haven't
curl -fsSL https://ollama.com/install.sh | sh

# Pull DeepSeek R1
ollama pull deepseek-r1:latest

Then configure Ishi to use Ollama:

{
  "provider": "ollama",
  "model": "deepseek-r1:latest"
}

Option 2: DeepSeek API

For cloud inference with DeepSeek's official API:

1. Get your API key from platform.deepseek.com
2. Configure in Ishi:

{
  "provider": "deepseek",
  "model": "deepseek-reasoner"
}

Glass Box: Seeing R1 Think

One of Ishi's core principles is Glass Box transparency. With R1's reasoning traces, you get even more visibility:

1. Draft — You describe your task
2. Reasoning Trace — Watch R1's chain-of-thought in real-time
3. Ghost Preview — See the proposed file changes
4. Execute — Approve and run

This is particularly powerful for complex tasks like:

• Analyzing financial documents
• Organizing large file collections
• Extracting structured data from PDFs

Real-World Example: Organizing 1,000 Downloads

Here's R1 reasoning through a file organization task:

"The user wants to organize their Downloads folder. Let me first categorize the file types: I see PDFs, images, zip files, and applications. The PDFs appear to be invoices based on filenames containing 'invoice' and dates. I'll create a folder structure: Documents/Invoices/2026-01/, Images/, Archives/, and Applications/. For the invoices, I'll extract the date from the filename and organize chronologically..."

This transparent reasoning lets you verify the logic before any files are moved.

Performance Notes

| Metric | DeepSeek R1 (Local) | DeepSeek R1 (API) | |--------|--------------------|--------------------| | First token | ~2-5s | ~1s | | Tokens/sec | 15-30 (M3 Mac) | 50+ | | Privacy | ✅ Full local | ⚠️ API calls | | Cost | Free | $0.14/1M input |

Get Started

1. Download Ishi if you haven't
2. Set up your preferred provider (Ollama or DeepSeek API)
3. Start with a simple task: "Organize my Desktop folder by file type"

The future of local AI is here. With R1 and Ishi, you get reasoning-level intelligence without sacrificing privacy.

---

Questions? Join our Discord or check the docs.