DQN-Labs/dqnGPT-v0.1-7B-MLX

🧠 DQN GPT v0.1 7B (MLX distribution)

High-Performance Local AI — No Datacenter Required.

DQN GPT 7B is a behavior-aligned, locally deployable assistant built on Mistral-7B-Instruct-v0.3, fine-tuned using an efficient QLoRA pipeline.

This release proves that powerful 7B models can be trained and deployed on modest hardware — sustainably, locally, and without enterprise infrastructure.

---

🚀 Mission

Make high-quality AI practical for everyday hardware.

DQN GPT is designed for:

• Students
• Indie developers
• Researchers
• Offline environments
• LAN-hosted assistants
• Personal AI servers

No subscriptions.
No cloud dependency.
No vendor lock-in.

---

🧠 Base Model

• Architecture: Mistral 7B Instruct v0.3
• Parameter Count: 7 Billion
• Context Length: 32,768 tokens
• Training Method: LoRA (low rank adaptation)
• Training Dataset: Simple pipline test (no major fine tune yet)
• Export Format: GGUF and MLX (llama.cpp / LM Studio compatible) (This MLX distribution of the model works only on LM Studio, and not on llama.cpp)

---

🔧 Fine-Tuning Approach

This model was fine-tuned using LoRA on consumer hardware (Apple M1).

Pipeline:

1. 4-bit quantized base model for memory efficiency
2. LoRA adapters trained on behavioral alignment data
3. Adapter fusion into full-precision base
4. Exported to FP16
5. Quantized to Q4_K_M for local deployment

Focus Areas:

• Identity consistency
• Stable conversational tone
• Reduced persona drift
• Clean instruction-following
• Practical, direct responses

This is not a benchmark-chasing release.
This is a usability-focused build.

---

💻 Hardware Requirements

Recommended (MLX 4-bit)

• 8GB RAM minimum , 16G RAM comfortable.
• CPU inference usable, even on low end devices.
• GPU optional, but speeds up inference significantly.

FP16 (for research / re-quantization)

• ~16GB RAM minimum, 32GB recommended.
• Intended for conversion, experimentation, or custom quantization.
• Very hard on CPU, GPU highly recommended.

FP16 is not available for download from the HF repo at the moment. If you would like access to the FP16 model, please contact me on Discord at @dqnlabs.

---

📦 Intended Use Cases

• Local AI assistant
• Offline productivity tool
• Personal coding helper
• Educational AI system
• Research experimentation
• LAN-hosted AI server
• Sustainable AI workflows

---

⚠ Limitations

• Early-stage release
• Behavior-focused fine-tune (not domain-specialized)
• Not optimized for coding benchmarks
• Not math-specialized
• Not tool-trained

This is a foundation release.

---

🛣 Roadmap

• Coding-specialized variant
• Hallucination reduction dataset
• Reasoning stability improvements
• Public evaluation benchmarks
• Larger context experimentation
• Structured instruction tuning

---

🌱 Sustainability Philosophy

Large models do not need large infrastructure.

With modern techniques like QLoRA, 7B models can:

• Be trained on modest hardware
• Be deployed locally
• Be distributed efficiently
• Reduce centralized compute dependence

AI should scale down — not just up.

---

🔓 License

Apache 2.0

Use it.
Modify it.
Deploy it.
Build on top of it.

---

🧭 DQN Labs

This model is part of the broader DQN Labs initiative:

Practical AI. Local-first. Engineer-minded.

More releases coming soon.