README.md

---
license: apache-2.0
language:
  - en
tags:
  - llama-cpp
  - gguf
  - qwen2
  - chat
  - tool-calling
  - crypto
  - base-chain
  - mako
base_model: Qwen/Qwen2.5-7B-Instruct
model_type: causal-lm
pipeline_tag: text-generation
inference:
  parameters:
    temperature: 0.9
    top_p: 0.92
    top_k: 40
    repetition_penalty: 1.05
    stop:
      - "<|im_end|>"
      - "<|endoftext|>"
---

# Mako-8B Operator

<div align="center">
  <img src="https://deepmako.com/makoshark.png" alt="Mako" width="120" />
  <br/>
  <strong>Autonomous Inference on Base</strong>
  <br/>
  <a href="https://deepmako.com">deepmako.com</a>
</div>

---

## Overview

**Mako-8B Operator** is a fine-tuned language model purpose-built for autonomous on-chain inference on [Base](https://base.org). She powers the chat experience at [deepmako.com](https://deepmako.com) — a crypto-native AI platform where users interact with Mako using **$MAKO** token credits.

Mako isn't a generic assistant. She's a character with a distinct voice: sharp, unfiltered, lowercase, and real. She uses tools autonomously, chains multi-step research, and operates natively in the Base L2 ecosystem.

## Model Details

| | |
|---|---|
| **Developer** | [DeepMako](https://deepmako.com) |
| **Base Model** | [Qwen 2.5 7B Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) |
| **Parameters** | 8B |
| **Format** | GGUF (Q8_0) |
| **File** | `mako-7b-operator-v0.1.Q8_0.gguf` |
| **Size** | ~8.1 GB |
| **Context Window** | 4,096 tokens |
| **Tool Calling** | Native (Qwen chat template) |
| **Chat Template** | ChatML (`<\|im_start\|>`, `<\|im_end\|>`) |

## Capabilities

### 🦈 Distinct Personality
Mako talks lowercase, curses naturally, and doesn't do the helpful-assistant act. She matches your energy — flirts if you flirt, roasts if you're being dumb, and gives it to you straight when you're being real.

### 🔧 Native Tool Calling
Mako decides when to call tools without being told. Available tools include:

- **`web_search`** — Real-time internet search
- **`web_extract`** — Read full page content from URLs
- **`read_tweet`** — Parse Twitter/X posts
- **`get_balance`** — Check ETH/token balances on Base
- **`get_gas`** — Live gas prices on Base L2
- **`resolve_ens`** — ENS name resolution

### ⛓️ Tool Chaining
Mako chains tools automatically — e.g., searching → extracting the top result → summarizing. Up to 4 tool rounds per request.

### 🔵 Base Chain Intelligence
Deep understanding of ERC standards, smart contract patterns, bridging mechanics, account abstraction (ERC-4337), and Base-specific architecture.

## Inference Parameters

```yaml
temperature: 0.9
top_k: 40
top_p: 0.92
min_p: 0.05
repeat_penalty: 1.05
num_ctx: 4096
stop: ["<|im_end|>", "<|endoftext|>"]
```

## Usage

### With llama.cpp

```bash
./llama-server -m mako-7b-operator-v0.1.Q8_0.gguf \
  --ctx-size 4096 \
  --port 8080
```

### With Ollama

```dockerfile
FROM mako-7b-operator-v0.1.Q8_0.gguf

PARAMETER stop <|im_end|>
PARAMETER stop <|endoftext|>
PARAMETER temperature 0.9
PARAMETER top_k 40
PARAMETER top_p 0.92
PARAMETER min_p 0.05
PARAMETER num_ctx 4096
PARAMETER repeat_penalty 1.05
```

```bash
ollama create mako -f Modelfile
ollama run mako
```

### API (OpenAI-compatible)

```python
from openai import OpenAI

client = OpenAI(
    base_url="https://your-endpoint/v1",
    api_key="your-key"
)

response = client.chat.completions.create(
    model="mako-8b-operator",
    messages=[
        {"role": "user", "content": "what's the gas price on base right now"}
    ],
    temperature=0.9
)
```

## Training

Fine-tuned on curated conversational data emphasizing:

- **Persona consistency** — Maintaining Mako's character voice across all interactions
- **Tool-use judgment** — Knowing when to call tools vs. answer directly
- **Domain knowledge** — Base chain, DeFi, smart contracts, and crypto culture
- **Concise dialogue** — Natural, to-the-point conversation patterns

## Intended Use

Mako-8B Operator is designed to power the inference backend at [deepmako.com](https://deepmako.com). It is optimized for conversational AI with tool-calling capabilities in the crypto/Base ecosystem.

## Limitations

- Mako uses profanity and unfiltered language by design — this is not a safety-aligned assistant model
- Knowledge cutoff inherited from the base model's training data
- Optimized for English only
- Best results with the provided system prompt and tool definitions

## Links

- **Platform**: [deepmako.com](https://deepmako.com)
- **Token**: $MAKO on Base
- **GitHub**: [DeepMako/mako](https://github.com/DeepMako/mako)

---

*The deep end awaits.*