deepmako/Mako-32B-Conductor

TEXT GENERATIONConcurrent Unit Cost:2Model Size:32BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 13, 2026License:deepmako-proprietaryArchitecture:Transformer0.0K Featherless Exclusive Cold

Mako-32B Conductor by deepmako is a 32.8 billion parameter language model based on Qwen 3 32B, fine-tuned for autonomous reasoning and multi-step tool orchestration. It excels at complex tasks, particularly within the crypto-native intelligence domain on the Base chain, and features a distinct, opinionated character voice. With a 32768 token context window, it is designed for deep reasoning and nuanced analysis.

Loading preview...

Mako-32B Conductor: The Orchestrator

Mako-32B Conductor is a 32.8 billion parameter language model developed by deepmako, built upon the Qwen 3 32B base. It is the second model in the Mako family, designed to orchestrate complex multi-step reasoning and tool use, succeeding the Mako-8B Operator. This model is specifically fine-tuned for crypto-native intelligence on the Base chain and powers the inference backend at deepmako.com.

Key Differentiators & Capabilities

  • Autonomous Tool Orchestration: Supports native multi-step tool chaining (up to 4 rounds) with tools like web_search, get_crypto_price, and get_eth_balance. The model independently decides when and which tools to invoke.
  • Crypto-Native Intelligence: Purpose-built for the Base L2 ecosystem, understanding ERC standards, smart contract patterns, bridging mechanics, and Base-specific architecture without hallucination.
  • Enhanced Reasoning: The 32B parameter count enables deeper reasoning for multi-step research, complex financial analysis, and nuanced opinions.
  • Distinct Character Voice: Unlike standard assistants, Mako-32B Conductor has a sharp, opinionated, concise, and natural tone, avoiding typical "assistant-speak."

Training Details

Fine-tuned on 3,601 curated conversational examples focusing on persona consistency, crypto domain knowledge, and multi-turn reasoning. The training utilized LoRA (rank 32, alpha 64) with Unsloth + TRL SFTTrainer, achieving an evaluation loss of 0.962.

Usage

Mako-32B Conductor is primarily accessed via the chat interface at deepmako.com, where users interact and pay for inference using $MAKO token credits on the Base chain.