Uluka Studio Mercury Mini is a fine-tuned Qwen3-8B GGUF model made for local assistant workflows, app-building help, and tool-oriented usage on consumer hardware.

It is designed to be small enough for local use while still being useful for structured prompts, planning, coding help, and lightweight agent-style workflows.

Details

Base model: Qwen/Qwen3-8B
Format: GGUF
Quantization: Q5_K_M
Fine-tune: Fireworks AI LoRA, manually merged
LoRA rank: 16
Training context length: 8192
Recommended prompt mode: /no_think for faster direct responses. Enable thinking only if your hardware can handle slower, longer outputs.

Use Cases

Uluka Studio Mercury Mini is intended for lightweight local assistant workflows where privacy, speed, and practical usefulness matter.

Good fits:

App-building help: planning features, improving UI ideas, generating small code snippets, and breaking projects into steps
Tool-oriented workflows: following structured prompts, producing clean outputs, and working inside simple local agent experiments
Coding support: explaining code, suggesting fixes, creating simple components, and organizing small projects
Local productivity: brainstorming, summarizing notes, drafting plans, and helping with everyday technical tasks
Experimentation: testing local AI workflows, prompt formats, and assistant systems without relying on a hosted API

Not ideal for:

High-stakes medical, legal, financial, or safety-critical advice
Complex software engineering without human review
Long autonomous agent runs where mistakes could cause real damage
Production use without additional testing and safeguards

This model works best as a small local assistant, not as a replacement for larger frontier models.

Recommended Specs

Minimum usable:

RAM: 12 GB system RAM
GPU: Optional, but recommended
Example hardware: CPU-only laptop or desktop with enough RAM
Use case: Short chats, light testing, lower context

Recommended:

RAM: 16 GB system RAM
GPU: GTX 1660 Super 6 GB, RTX 2060 6 GB, RTX 3050 6 GB, or similar
Use case: Simple local chat, coding help, and light agentic use with partial GPU offload

Best experience:

RAM: 32 GB system RAM
GPU: 8 GB+ VRAM
Example hardware: RTX 3060 12 GB, RTX 4060 Ti 16 GB, RTX 4070, or better
Use case: Better speed, longer chats, higher context, and smoother local assistant workflows

LM Studio

This model should work in LM Studio as a local GGUF model. For private repos, download the .gguf file manually and import it into LM Studio.

Usage

Example llama.cpp command:

bash

llama-cli \
  -m Uluka-Studio-Mercury-Mini-8B-Q5_K_M.gguf \
  -p "<|im_start|>user\n/no_think Help me plan a simple app builder UI.<|im_end|>\n<|im_start|>assistant\n" \
  -n 300 \
  --temp 0.7

License / Use Terms

This release is provided for personal use only.

You may use this model for private testing, learning, experimentation, and personal local projects.

You may not sell, host, resell, redistribute, re-upload, or use this model as part of a paid product or public service without permission.

This model is based on Qwen/Qwen3-8B. Users must also follow the base model license and any applicable terms from the fine-tuning data.

UlukaDev/uluka-studio-mercury-mini-8b-q5-k-m

Uluka Studio Mercury Mini 8B Q5_K_M