mango-1.0

mango-1.0 (Q2_K_L GGUF)

Local coding / agent model for Mango — a desktop coding agent built around small GGUF runtimes (llama.cpp). This repo ships a quantized GGUF of Mango 1.0 (~27B, Qwen3.8 family), fine-tuned for tool-using agent workflows (short thoughts, tool calls, file edits, Q&A over a workspace).

Model details

Intended use

• Run inside Mango (Electron + Python sidecar + llama.cpp), or any llama.cpp-compatible runner (LM Studio, koboldcpp, etc.).
• Coding assistance: read/search/edit files, answer questions about a workspace, follow agent tool protocols.
• Offline / privacy-sensitive development.

Not intended for

• High-stakes medical, legal, or financial advice
• Unsupervised production systems without human review
• Tasks that need full FP16 quality (this is a heavy 2-bit-class quant)

How to use

Mango app

1. Download mango-1.0-Q2_K_L.gguf
2. In Mango → Settings, set the local .gguf path
3. Start a session and use /ask, /plan, or agent mode as usual

llama.cpp / LM Studio

Load the GGUF like any other Qwen ChatML model. Prefer a GPU offload that fits your VRAM; leave layers on CPU if needed. Example (llama.cpp CLI sketch):

./llama-cli -m mango-1.0-Q2_K_L.gguf -c 8192 -ngl 99 -p "You are a coding assistant."

Training (high level)

• Base: Qwen3.8-class 27B (Unsloth / QLoRA pipeline in the Mango training scripts)
• Method: 4-bit QLoRA → merge → GGUF export
• Focus: Agent/coding-style SFT (tool use, short CoT, workspace workflows) Exact datasets and hyperparameters live in the Mango public repo under training/.

Limitations

• Quantization: Q2_K_L is aggressive — expect weaker long-horizon reasoning and more tool-format slips than higher quants (Q4/Q5/Q8) or FP16.
• Hallucinations: May invent APIs/paths if not grounded by tools; prefer Ask/Agent modes that read files.
• Safety: Same class of risks as other open coding models (unsafe code suggestions, prompt injection via repo content).

License

Apache-2.0 (unless otherwise noted for the base model / training data — check base model cards as well).

Citation

@misc{mango10_gguf,
  title        = {Mango 1.0 GGUF},
  author       = {ToldByNun},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/ToldByNun/mango-1.0-iq2-xs}},
  note         = {Qwen3.8-based local coding agent model, Q2\_K\_L GGUF}
}

Links

• App / agent: github.com/ToldByNun/mango-public
• Model: huggingface.co/ToldByNun/mango-1.0-iq2-xs