---
license: apache-2.0
language:
- en
tags:
- mlx
- distill
- cli
- code
- compression
- qwen
- expert-model
- domain-specific
- task-specialized
pipeline_tag: text-generation
base_model: Qwen/Qwen3-1.7B
---

# distill-1.7B — Expert Language Model for CLI Output

**distill-1.7B** is a **domain-specific Expert Language Model** — not a general-purpose chatbot. It does exactly one thing: compress and classify raw terminal output into structured, actionable summaries.

Built for the **[distill](https://github.com/samuelfaj/distill)** engine — an open-source CLI output compression tool.

## What is distill?

[distill](https://github.com/samuelfaj/distill) is a tool that takes arbitrary command-line output and reduces it to only what matters. Instead of scrolling through 500 lines of `npm install` logs, you get:

```
PASS
24 packages installed, 0 vulnerabilities
```

Instead of parsing a wall of Terraform plan output, you get:

```json
{"create": 3, "change": 12, "destroy": 0}
```

**distill-1.7B is the brain behind distill** — it's the model that understands CLI output and knows what's signal vs noise.

## Why "Expert Language Model"?

Unlike general-purpose LLMs (ChatGPT, Claude, etc.) that can talk about anything, distill-1.7B is:

| Trait | General LLM | distill-1.7B |
|-------|-------------|--------------|
| **Scope** | Any topic | CLI output only |
| **Size** | 70-400B params | 1.7B params |
| **Training data** | Web crawl (trillions of tokens) | 100k synthetic CLI outputs |
| **Strengths** | Conversation, reasoning, code | CLI compression, classification |
| **Weaknesses** | — | Can't chat, can't code, can't reason |

It's an **expert** in the same way a radiologist is an expert — highly skilled in one narrow domain, not trying to be a general practitioner.

## 8 Specialized Tasks

| Task | What it does | Example output |
|------|-------------|----------------|
| `pass_fail` | Did the command succeed or fail? | `PASS` / `FAIL Error: ...` |
| `safe_review` | Is this Terraform plan safe? | `SAFE` / `UNSAFE` / `REVIEW` |
| `terraform_plan` | Count resources created/changed/destroyed | `{"create":3,"change":12,"destroy":0}` |
| `json_extraction` | Pull JSON from noisy logs | `[{"name":"app","version":"2.1.0"}]` |
| `security_audit` | Count vulns by severity | `[{"severity":"high","count":2}]` |
| `test_result` | Test suite pass/fail? | `PASS\n4 passed, 0 failed` |
| `typescript_check` | Extract TS compiler errors | `error TS2741: Property 'x' is missing` |
| `generic` | Free-form summary of any CLI output | `24 packages installed` |

## Performance

| Metric | Value |
|--------|-------|
| Overall accuracy | **95%** |
| Tasks at 100% | 6 of 8 |
| Base model | Qwen3-1.7B |
| Training | LoRA rank 32, 4000 iterations |
| Dataset | 100k synthetic CLI outputs |
| Training hardware | Apple M4 Max, 128 GB RAM |

## Available Formats

| Repo | Format | Size | Platform |
|------|--------|------|----------|
| **distill-1.7B-MLX** | MLX fp16 | 3.2 GB | macOS (Apple Silicon) |
| [distill-1.7B-4bit-MLX](https://huggingface.co/samuelfaj/distill-1.7B-4bit-MLX) | MLX 4-bit | 1.0 GB | macOS (Apple Silicon) |
| [distill-1.7B-GGUF](https://huggingface.co/samuelfaj/distill-1.7B-GGUF) | GGUF fp16 | 4.1 GB | Cross-platform |
| [distill-1.7B-4bit-GGUF](https://huggingface.co/samuelfaj/distill-1.7B-4bit-GGUF) | GGUF Q4_K_M | 1.2 GB | Cross-platform |

All formats achieve **identical 95% accuracy** — pick based on your platform and size preference.

## Usage

```python
from mlx_lm import load, generate

model, tokenizer = load("samuelfaj/distill-1.7B-MLX")

messages = [
    {"role": "system", "content": "You are distill. Compress CLI output concisely."},
    {"role": "user", "content": "Command output:\nnpm test\n4 tests passed, 0 failed"}
]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
result = generate(model, tokenizer, prompt=prompt, max_tokens=256)
print(result)
```

## Project

This model powers [distill](https://github.com/samuelfaj/distill) — a CLI output compression engine. The training code and dataset generation pipeline are available in the repository.

[Full Distill Collection](https://huggingface.co/collections/samuelfaj/distill-6a0606f9b131c289025659fc)