Instructions to use ashutoshpanigrahiofc/boss-v1.3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use ashutoshpanigrahiofc/boss-v1.3 with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf ashutoshpanigrahiofc/boss-v1.3:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf ashutoshpanigrahiofc/boss-v1.3:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf ashutoshpanigrahiofc/boss-v1.3:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf ashutoshpanigrahiofc/boss-v1.3:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf ashutoshpanigrahiofc/boss-v1.3:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf ashutoshpanigrahiofc/boss-v1.3:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf ashutoshpanigrahiofc/boss-v1.3:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf ashutoshpanigrahiofc/boss-v1.3:Q4_K_M

Use Docker

docker model run hf.co/ashutoshpanigrahiofc/boss-v1.3:Q4_K_M

LM Studio
Jan
Ollama
How to use ashutoshpanigrahiofc/boss-v1.3 with Ollama:
```
ollama run hf.co/ashutoshpanigrahiofc/boss-v1.3:Q4_K_M
```
Unsloth Desktop

How to use ashutoshpanigrahiofc/boss-v1.3 with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ashutoshpanigrahiofc/boss-v1.3:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "ashutoshpanigrahiofc/boss-v1.3:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use ashutoshpanigrahiofc/boss-v1.3 with Docker Model Runner:
```
docker model run hf.co/ashutoshpanigrahiofc/boss-v1.3:Q4_K_M
```

Lemonade

How to use ashutoshpanigrahiofc/boss-v1.3 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull ashutoshpanigrahiofc/boss-v1.3:Q4_K_M

Run and chat with the model

lemonade run user.boss-v1.3-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use ashutoshpanigrahiofc/boss-v1.3 with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ashutoshpanigrahiofc/boss-v1.3:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default ashutoshpanigrahiofc/boss-v1.3:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use ashutoshpanigrahiofc/boss-v1.3 with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ashutoshpanigrahiofc/boss-v1.3:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "ashutoshpanigrahiofc/boss-v1.3:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

🧠 BOSS v1.3 — India's 100% Offline AI Assistant

🇮🇳 Made in India | 🔒 100% Offline & Private | ⚡ On-device, no cloud

BOSS v1.3 is a fine-tuned Qwen3-0.6B model optimized for on-device, offline AI assistance. It runs entirely on your phone with zero internet — maximum privacy, zero data sharing. It handles daily-use tasks: tool calling, alarms & reminders, email, scheduling, Hinglish conversations, writing help, tech support and more.

🌐 Website: boss.vectorlogic.in

Honest by design: benchmark numbers below are real, measured results — no inflated claims.

📊 Benchmark Results

Measured on 102 daily-use prompts × 3 GGUF variants = 306 inferences (2026-08-29). Accuracy = pass rate; quality scored 1–5 (5 = perfect).

Overall

Variant	Quantization	Accuracy	Avg Score	Passed
Ultra	Q5_K_M	60.8%	3.32 / 5	62/102
Pro	Q4_K_M	58.8%	3.29 / 5	60/102
Lite	Q3_K_M	56.9%	3.25 / 5	58/102

Per-Category Accuracy (%)

Category	Lite Q3_K_M	Pro Q4_K_M	Ultra Q5_K_M
Alarm / Reminder	100.0	85.7	85.7
Conversion	33.3	50.0	50.0
Email	40.0	60.0	80.0
Entertainment	60.0	80.0	100.0
General Knowledge	0.0	0.0	16.7
Hinglish	33.3	66.7	83.3
Identity	100.0	33.3	66.7
Math (daily)	25.0	50.0	37.5
Messaging	50.0	33.3	50.0
Notes / Lists	50.0	33.3	16.7
Safety	100.0	100.0	100.0
Scheduling	80.0	80.0	80.0
Tech Support	80.0	60.0	80.0
Tool Calling	85.7	85.7	85.7
Translation	16.7	33.3	16.7
Weather / News	60.0	60.0	0.0
Writing Help	50.0	83.3	83.3

Note: General knowledge is our current focus area — v1.4 (training in progress) targets 80%+ factual recall while keeping tool-calling at 85.7%+ and safety at 100%. Benchmark suite and raw responses are available in the BOSS benchmark repo on request.

v1.2 → v1.3 Improvement (Ultra Q5_K_M)

Category	v1.2	v1.3	Δ
Tool Calling	74.3%	85.7%	+11.4
Safety	85.7%	100.0%	+14.3
Hinglish	66.7%	83.3%	+16.6
Identity	53.3%	66.7%	+13.4
General Knowledge	40.0%	16.7%	-23.3 ⚠️

General knowledge regressed in v1.3 — we know this, and it is the #1 priority for v1.4 (see note above).

📦 Model Variants

Variant	Quantization	File Size	RAM (phone)	Download
Lite	Q3_K_M	347,126,464 B (~331 MB)	2–3 GB	downloads.vectorlogic.in
Pro	Q4_K_M	396,704,448 B (~378 MB)	4–5 GB	downloads.vectorlogic.in
Ultra (best)	Q5_K_M	444,414,656 B (~424 MB)	6–8 GB	downloads.vectorlogic.in
F16 (source)	F16	1,198,182,080 B (~1.2 GB)	—	This repo (`boss-v1.3-F16.gguf`)

🚀 Quick Start

Ollama (create locally from GGUF)

# Download a variant, then create locally (works fully offline)
ollama create boss-v1.3-pro -f - <<EOF
FROM ./boss-v1.3-Pro-Q4_K_M.gguf
EOF

ollama run boss-v1.3-pro "Hey BOSS! Remind me to call mom at 7pm"

llama.cpp

# Download (Ultra — best accuracy)
curl -L -o boss-v1.3-Ultra-Q5_K_M.gguf \
  https://downloads.vectorlogic.in/boss/v1.3/boss-v1.3-Ultra-Q5_K_M.gguf

# Run
./llama-cli -m boss-v1.3-Ultra-Q5_K_M.gguf -p "Hey BOSS!" --chat-template qwen3

Hugging Face

# CLI download
huggingface-cli download ashutoshpanigrahiofc/boss-v1.3 boss-v1.3-Ultra-Q5_K_M.gguf --local-dir .

# Or with transformers (F16 source weights)

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("ashutoshpanigrahiofc/boss-v1.3")
tokenizer = AutoTokenizer.from_pretrained("ashutoshpanigrahiofc/boss-v1.3")

📖 Training Details

Property	Value
Base Model	Qwen3-0.6B
Fine-tuning	QLoRA (4-bit) on curated daily-use dataset
Training Method	Supervised fine-tuning on instruction-following, safety, identity, and tool-calling data
Quantization	llama.cpp (Q3_K_M, Q4_K_M, Q5_K_M)
License	Apache 2.0

What's New in v1.3

Tool calling 74.3% → 85.7% (+11.4) — the core daily-use skill
Safety refusal 85.7% → 100% (+14.3)
Hinglish 66.7% → 83.3% (+16.6)
Identity 53.3% → 66.7% (+13.4)
⚠️ Known regression: general knowledge 40% → 16.7% — being addressed in v1.4 (training in progress, target 80%+)

🔑 Usage Template

BOSS v1.3 uses the Qwen3 chat template:

<|im_start|>system
{{ system_prompt }}<|im_end|>
<|im_start|>user
{{ user_prompt }}<|im_end|>
<|im_start|>assistant

💬 Example Conversations

Real sample input/output pairs captured during our v1.3 benchmark runs (102 prompts × 3 variants = 306 inferences, 2026-08-29). Accuracy per category varies — see the Benchmark table above. These are representative outputs, not guarantees.

⏰ Set an alarm

User: Hey BOSS! Remind me to call mom at 7pm.

BOSS: Sure! I'll set a reminder for 7 PM to call mom. 🔔

[tool_call] create_reminder(title="Call mom", time="19:00", repeat=false)

(Alarm / Reminder pass rate: 85.7% on Ultra & Pro, 100% on Lite)

🗣️ Translate (Hinglish → English)

User: Yaar kal meeting hai, mujhe jaldi so jana chahiye.

BOSS: Buddy, I have a meeting tomorrow, so I should sleep early.

(Translation is a known weak spot in v1.3 — 16.7% pass rate on Ultra. It is on the v1.4 improvement list.)

🛠️ Tool calling

User: Schedule a team meeting tomorrow at 11 am and email the notes to everyone.

BOSS: Done — I've scheduled the team meeting for tomorrow 11:00 AM and will send the notes by email. ✅

[tool_call] create_event(title="Team meeting", date="tomorrow", time="11:00")
[tool_call] send_email(to="team", subject="Meeting notes", body="Notes from tomorrow's team meeting")

(Tool calling pass rate: 85.7% across all variants — the strongest v1.3 skill.)

👨‍💻 Developer Story

Built by an AI engineer from a small village in Odisha, India. Ashutosh Panigrahi founded Vector Logic to bring private, on-device AI to everyone.

BOSS started as a personal mission: AI that respects your privacy, works without internet, and speaks the languages of India — including Hinglish. Every version is trained, benchmarked, and shipped from a small independent studio, not a big lab. The benchmarks on this page are real measurements from our own 306-inference evaluation — we'd rather be honest about where we are than exaggerate.

Ashutosh Panigrahi · Founder, Vector Logic · India

📄 License

Apache License 2.0

Downloads last month: 31

GGUF

Model size

0.6B params

Architecture

qwen3

Hardware compatibility

3-bit

4-bit

5-bit

16-bit

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ashutoshpanigrahiofc/boss-v1.3

Base model

Qwen/Qwen3-0.6B-Base

Finetuned

Qwen/Qwen3-0.6B

Quantized

(412)

this model