---
library_name: mlx
license: apache-2.0
license_link: https://huggingface.co/Qwen/Qwen3-0.6B/blob/main/LICENSE
pipeline_tag: text-generation
base_model: Qwen/Qwen3-0.6B
tags:
- mlx
- safetensors
- qwen3
- fine-tuned
- consulting
- routing
- tool-calling
- edge
- apple-silicon
- on-device
language:
- en
model-index:
- name: analyst-0.6b
  results: []
---

# Analyst 0.6B

A fine-tuned [Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B) specialist for consulting-domain AI workflows. Built by [AXe Technologies](https://axe.onl) for production deployment in the [Pulse](https://consultimi.com) platform.

## Overview

Analyst 0.6B is a domain-tuned small language model designed for **fast routing, intent classification, and structured call construction** in consulting and professional services contexts. It runs entirely on-device — Apple Silicon Macs, edge servers, or any hardware that supports MLX or GGUF inference.

| Spec | Value |
|------|-------|
| Parameters | 0.6B |
| Base Model | Qwen3-0.6B |
| Format | MLX (safetensors) |
| Training | LoRA fine-tune, single epoch |
| Context | 32K tokens |
| License | Apache 2.0 |

## Intended Use

- **Intent routing** — classify user turns and dispatch to appropriate specialist models
- **Call construction** — parse natural language into structured function calls
- **Domain drafting** — generate consulting-domain responses with professional tone
- **SQL generation** — natural language to SQL for business analytics (basic queries)

Designed as the fast first-pass router in a multi-model specialist pipeline. Pairs well with larger models (3B, 7B) for complex reasoning tasks.

## Quickstart

### MLX (Apple Silicon)

```python
from mlx_lm import load, generate

model, tokenizer = load("axetechnologies/analyst-0.6b")
prompt = "Classify this consulting request: 'Show me revenue by region for Q3'"
response = generate(model, tokenizer, prompt=prompt, max_tokens=256)
print(response)
```

### llama.cpp / Ollama

Convert to GGUF for cross-platform inference:

```bash
# Using mlx_lm to convert, or download GGUF variants when available
python -m mlx_lm.convert --hf-path axetechnologies/analyst-0.6b --quantize q8_0
```

## Training

- **Method:** LoRA (r=16, 16 target layers, alpha=32)
- **Learning rate:** 1e-4
- **Batch size:** 2-4
- **Iterations:** 400
- **Epochs:** 1 (single epoch — multi-epoch degrades instruction-tuned bases)
- **Hardware:** Apple Silicon (Mac Studio M2 Ultra, 64GB)
- **Framework:** MLX with mlx-lm

Training data is a curated mix of consulting-domain interactions: routing decisions, methodology checks, narrative interpretation, and NL-to-SQL pairs.

## Limitations

- Optimized for consulting/professional services domain — general-purpose performance may trail the base model on out-of-domain tasks
- 0.6B parameter count means complex multi-step reasoning should be delegated to larger specialists
- English only

## Model Family

| Model | Parameters | Role | Status |
|-------|-----------|------|--------|
| **analyst-0.6b** | 0.6B | Router / fast classifier | Released |
| [analyst-3b](https://huggingface.co/axetechnologies/analyst-3b) | 3B | Call construction / parsing | Released |
| [analyst-7b](https://huggingface.co/axetechnologies/analyst-7b) | 7B | Drafting / narrative | Released |

## About

Built by [AXe Technologies](https://axe.onl) — sovereign AI infrastructure for regulated industries. All training and inference runs on owned hardware in Canada. No data leaves the perimeter.