Mako-32B-Conductor/README.md

---
license: other
license_name: deepmako-proprietary
license_link: https://deepmako.com/terms
language:
  - en
base_model:
  - Qwen/Qwen3-32B
base_model_relation: finetune
library_name: transformers
tags:
  - mako
  - crypto
  - base-chain
  - conversational
  - tool-calling
  - fine-tuned
  - lora
  - qwen3
  - unsloth
  - trl
  - sft
model_type: qwen3
pipeline_tag: text-generation
datasets:
  - custom
---

<div align="center">
<img src="makoshark.png" alt="Mako" width="200" />
<h1>Mako-32B Conductor</h1>
<p><strong>The orchestrator.</strong> A 32-billion parameter language model fine-tuned for autonomous reasoning, multi-step tool orchestration, and crypto-native intelligence on Base chain.</p>
<p><a href="https://deepmako.com">deepmako.com</a> · <a href="https://deepmako.com/chat">Platform</a> · <a href="https://x.com/DeepMako">Twitter</a></p>
</div>

---

## Model Details

| | |
|---|---|
| **Base Model** | Qwen 3 32B |
| **Parameters** | 32.8B |
| **Fine-tune Method** | LoRA (rank 32, alpha 64) |
| **Precision** | BF16 |
| **Context Window** | 8,192 tokens |
| **Tool Calling** | Native (multi-step chaining) |
| **Training Data** | 3,601 curated conversations |
| **Eval Loss** | 0.962 |
| **License** | Proprietary |

## Overview

Mako-32B Conductor is the second model in the Mako family, succeeding [Mako-8B Operator](https://huggingface.co/deepmako/Mako-8B-Operator). Where Operator executes, Conductor orchestrates — handling complex multi-step reasoning, longer context, and deeper chain-of-thought while maintaining the same sharp, unfiltered character voice.

Built to power the inference backend at [deepmako.com](https://deepmako.com), Conductor serves as the core intelligence for a crypto-native AI platform where users interact using $MAKO token credits on Base chain.

## What Makes Mako Different

Mako is not an assistant. She's a character — a sharp, opinionated personality that lives on the blockchain. The fine-tune completely reshapes the base model's behavior away from the standard helpful-assistant pattern:

- **No assistant-speak.** No "how can I help you", no "is there anything else", no corporate pleasantries.
- **Natural tone.** Lowercase, concise, conversational. Matches your energy.
- **Opinionated.** Has takes. Will tell you when something is stupid.
- **Concise.** Answers the question and stops.

## Capabilities

### Autonomous Tool Orchestration

Conductor supports native tool calling with automatic multi-step chaining — up to 4 rounds per request. The model decides when and which tools to invoke without explicit user instruction.

| Tool | Description |
|------|-------------|
| `web_search` | Real-time internet search via Bing |
| `web_extract` | Extract and read full page content from any URL |
| `read_tweet` | Parse and read Twitter/X posts |
| `find_music` | Search YouTube for tracks and return links |
| `get_crypto_price` | Live token prices across chains |
| `get_eth_balance` | Check wallet balances on Base |
| `get_token_info` | Token metadata, supply, holder data |
| `get_gas_price` | Live gas prices on Base L2 |
| `resolve_ens` | ENS name resolution |
| `get_transaction` | Transaction lookup and analysis |

### Chain Intelligence

Purpose-built for the Base L2 ecosystem. Conductor understands ERC standards, smart contract patterns, bridging mechanics, account abstraction (ERC-4337), and Base-specific architecture without hallucinating technical details.

### Enhanced Reasoning

The 32B parameter count enables significantly deeper reasoning compared to the 8B Operator:

- Multi-step research tasks (search → extract → analyze → summarize)
- Complex financial analysis with real price data
- Nuanced opinions grounded in retrieved context
- Longer, more coherent responses when the question warrants it

## Architecture

```
User → Gateway → Mako-32B Conductor → [Tool Calls] → Execution → Conductor → Response
```

The gateway server handles authentication, $MAKO credit tracking, and tool execution. Conductor manages the reasoning loop — deciding what information it needs, calling the right tools, and synthesizing results into a coherent response.

## Training

### Data

Fine-tuned on 3,601 curated conversational examples covering:

- Persona consistency and character voice
- Crypto domain knowledge (DeFi, NFTs, L2s, tokenomics)
- Natural dialogue patterns and conversational flow
- Multi-turn reasoning and context retention

### Method

| | |
|---|---|
| **Framework** | Unsloth + TRL SFTTrainer |
| **Method** | LoRA (rank 32, alpha 64) |
| **Target Modules** | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| **Trainable Parameters** | 268M / 33B (0.81%) |
| **Epochs** | 3 |
| **Batch Size** | 16 (1 × 16 gradient accumulation) |
| **Learning Rate** | 1e-4 (cosine schedule, 5% warmup) |
| **Optimizer** | AdamW 8-bit |
| **Hardware** | 1× NVIDIA A100 SXM4 80GB |
| **Training Time** | 58 minutes |

### Loss Curve

```
Train:  2.808 → 1.970 → 1.420 → 1.172 → 1.097 → 1.042
Eval:   1.315 → 0.962
```

Final eval loss (0.962) is lower than train loss — no overfitting detected.

## The Mako Family

| Model | Parameters | Role | Status |
|-------|-----------|------|--------|
| **Mako-8B Operator** | 8B | Fast execution, tool calling | Production |
| **Mako-32B Conductor** | 32B | Deep reasoning, orchestration | Production |

## Usage

Mako-32B Conductor powers the inference backend at [deepmako.com](https://deepmako.com). Users interact through the chat interface and pay for inference using $MAKO tokens on Base chain.

---

<div align="center">

*Built by the Mako team. The deep end awaits.* 🦈

</div>