Instructions to use DJLougen/Harmonic-27B-MLX-8bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use DJLougen/Harmonic-27B-MLX-8bit with MLX:

# Make sure mlx-lm is installed
# pip install --upgrade mlx-lm

# Generate text with mlx-lm
from mlx_lm import load, generate

model, tokenizer = load("DJLougen/Harmonic-27B-MLX-8bit")

prompt = "Write a story about Einstein"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True
)

text = generate(model, tokenizer, prompt=prompt, verbose=True)

Notebooks
Google Colab
Kaggle
Local Apps
LM Studio

Unsloth Studio new

How to use DJLougen/Harmonic-27B-MLX-8bit with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for DJLougen/Harmonic-27B-MLX-8bit to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for DJLougen/Harmonic-27B-MLX-8bit to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for DJLougen/Harmonic-27B-MLX-8bit to start chatting

Load model with FastModel

pip install unsloth
from unsloth import FastModel
model, tokenizer = FastModel.from_pretrained(
    model_name="DJLougen/Harmonic-27B-MLX-8bit",
    max_seq_length=2048,
)

Pi new

How to use DJLougen/Harmonic-27B-MLX-8bit with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "DJLougen/Harmonic-27B-MLX-8bit"

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "DJLougen/Harmonic-27B-MLX-8bit"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use DJLougen/Harmonic-27B-MLX-8bit with Hermes Agent:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "DJLougen/Harmonic-27B-MLX-8bit"

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default DJLougen/Harmonic-27B-MLX-8bit

Run Hermes

hermes

MLX LM

How to use DJLougen/Harmonic-27B-MLX-8bit with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "DJLougen/Harmonic-27B-MLX-8bit"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "DJLougen/Harmonic-27B-MLX-8bit"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "DJLougen/Harmonic-27B-MLX-8bit",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Harmonic-27B-MLX-8bit

MLX 8-bit quantized conversion of DJLougen/Harmonic-27B — the flagship of the Harmonic series. A reasoning-focused fine-tune of Qwen 3.5 27B trained on structurally validated data. Every row passes automated quality gates. No junk, no filler, no shallow traces.

The name comes from harmonic analysis of reasoning patterns — the structural signal that separates genuine thinking from surface-level chain-of-thought.

Support This Work

I'm a PhD student in visual neuroscience at the University of Toronto who also happens to spend way too much time fine-tuning, merging, and quantizing open-weight models on rented H100s and a local DGX Spark. All training compute is self-funded — balancing GPU costs against a student budget. If my uploads have been useful to you, consider buying a PhD student a coffee. It goes a long way toward keeping these experiments running.

Support on Ko-fi

Quantization Details

Format: MLX SafeTensors
Bits per weight: 8.50
Note: 8-bit quantization — near-lossless quality. Requires ~27GB.

Training Approach

Same pipeline as Harmonic-9B. 799 curated rows — a small, precisely curated dataset instead of tens of thousands of unfiltered examples. The base model already has the knowledge from pretraining — the fine-tune teaches it a reasoning behavior pattern.

Every training row contains explicit self-correction ("wait, that's not right"), verification ("let me check by plugging back in"), and multi-path exploration ("alternatively, I could try..."). The data was generated from multiple frontier models and filtered through a custom structural quality pipeline that enforces reasoning depth, coherence, and flow patterns. 100% of rows pass all quality gates simultaneously.

Training Data Quality

The same reasoning data as Harmonic-9B and Harmonic-2B, curated using a custom structural process supervision pipeline:

Metric	Value
Signal quality score	78.7 mean (61.5 min, 90.0 max)
Thinking trace depth	1,667 words average
Self-correction	100% of rows (17.2 per row avg)
Verification	100% of rows (10.3 per row avg)
Exploration	100% of rows (6.3 per row avg)
Quality gate pass rate	100%

How It Compares

We ran our structural quality analysis against every major public reasoning dataset used for Opus/Qwen distillation. The results:

Dataset	Rows	Think Words	Self-Correction	Verification	Exploration	Signal Score	Gate Pass
Harmonic (ours)	799	1,667	100%	100%	100%	78.7	100%
Crownelius/Opus-3300x	2,160	188	5.9%	22.6%	5.2%	28.0	0.1%
nohurry/Opus-Filtered	2,326	191	6.7%	24.1%	5.3%	28.5	0.1%
TeichAI/Opus-250x	250	323	17.2%	26.8%	6.8%	24.6	0.4%
Jackrong/Qwen-700x	633	6,653	97.5%	97.6%	69.8%	75.6	22.7%
Bespoke-Stratos-17k	16,710	1,322	88.2%	72.7%	59.7%	71.7	49.0%
glaiveai/reasoning-20m	22M+	799	64.1%	41.4%	37.3%	46.2	12.8%
KingNish/reasoning-20k	19,944	132	0.7%	4.2%	4.3%	27.4	0.0%

Speculative Decoding

Harmonic-27B pairs with Harmonic-2B for speculative decoding. Both models share the same training data, reasoning format, and architecture family (Qwen 3.5), which keeps draft token acceptance rates high.

from transformers import AutoModelForCausalLM

target = AutoModelForCausalLM.from_pretrained("DJLougen/Harmonic-27B")
draft = AutoModelForCausalLM.from_pretrained("DJLougen/Harmonic-2B")

outputs = target.generate(
    **inputs,
    assistant_model=draft,
    max_new_tokens=512,
)

Usage

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("DJLougen/Harmonic-27B-MLX-8bit")
response = generate(model, tokenizer, prompt="What is the sum of the first 100 prime numbers?", max_tokens=2048)
print(response)

Reasoning format

The model uses think blocks for reasoning:

<|thinking|>
The user is asking about X. Let me consider two approaches...

Approach 1: ...
Approach 2: ...

I will go with Approach 1 because...

Wait, I need to be careful here - this assumes Y, which may not hold.
Let me verify by checking a special case...

Yes, that confirms the result.
<|/thinking|>

[Final answer here]

Training Configuration

base_model: unsloth/Qwen3.5-27B
dataset: 799 curated reasoning rows
epochs: 1
learning_rate: 1e-4
lr_scheduler: cosine
warmup_ratio: 0.1
max_seq_length: 8192
lora_rank: 32
lora_alpha: 32
dropout: 0.05
micro_batch_size: 1
gradient_accumulation_steps: 4
weight_decay: 0.01

Intended Use

Reasoning tasks requiring genuine multi-step thinking
Mathematical problem-solving with self-correction
Code analysis and generation with structured verification
General conversation (conversational ability preserved through training design)
Target model for speculative decoding with Harmonic-2B
Base model for Stage 2 agentic fine-tuning

Limitations

Reasoning traces can be verbose for simple questions
Not optimized for tool calling — see Harmonic-Hermes-9B for agentic use
Benchmark evaluation is ongoing

Architecture

Base: Qwen 3.5 27B (27.36B parameters)
Training: LoRA fine-tuning, merged into base weights
Original Precision: BF16
Context: 8192 tokens

License

Apache 2.0 — same as the base model. All training data is from Apache 2.0 or MIT licensed sources. Fully commercial use permitted.

Model tree for DJLougen/Harmonic-27B-MLX-8bit

Base model

Qwen/Qwen3.5-27B

Finetuned

unsloth/Qwen3.5-27B