Bragi-LLM-GGUF
An 805 MB local Python coding assistant. 92% MBPP single-shot, 2 points behind Qwen2.5-Coder-7B (14 GB, 17x larger). Zero API cost.
What is this
c15v-q3km-imat.gguf is the backbone GGUF for the Bragi-LLM system, a 786 MB Q3_K_M quantised Qwen2.5-Coder-1.5B-Instruct with imatrix calibration on MBPP-train Python code. It is the small LLM half of a hybrid design: combined with a 15 KB hand-engineered symbolic engine library and a 6 KB keyword intercept router, the full system reaches 92% pass@1 on the MBPP test split, single-shot greedy decoding, no retry. See the paper for the failure-mode analysis and ablations: doi:10.5281/zenodo.20557449.
The triptych
| Role | Repo | What it is |
|---|---|---|
| Brain (this file) | Bragi-LLM | The 805 MB local Python coder. |
| Eyes | Code Tree | The visual IDE. |
| Hands | Demeter-CodeBuilder | OpenAI-compatible proxy wiring Bragi as Code Tree's default local backend. |
Together: about 1 GB on disk. Fully offline. Zero recurring fees. MIT.
Results
| System | Footprint | MBPP test 0-99 single-shot |
|---|---|---|
| Vanilla 1.5B Q3_K_M (this backbone alone) | 786 MB | 65% |
| With intercept router + engine_lib (full Bragi system) | 805 MB | 92% |
| Reference Qwen2.5-Coder-7B fp16 | 14 GB | 94% |
The GGUF alone (this file) is the 65% column. The 92% number requires the router and engine library which are open-sourced in the Bragi-LLM GitHub repo.
Quick start
Download
huggingface-cli download norika1207-lab/Bragi-LLM-GGUF c15v-q3km-imat.gguf --local-dir .
Run with llama.cpp
# build llama.cpp once
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp && cmake -B build && cmake --build build --config Release
# serve
./build/bin/llama-server -m c15v-q3km-imat.gguf -ngl 99 -c 16384 --parallel 4 --port 8080
# CPU only
./build/bin/llama-server -m c15v-q3km-imat.gguf -ngl 0 -c 16384 --parallel 4 --port 8080
Use full Bragi system (recommended, gets 92%)
Combine with router and engine library from the GitHub repo:
git clone https://github.com/norika1207-lab/Bragi-LLM
cd Bragi-LLM
python3 solve_intercept2.py 1 1 100 http://localhost:8080/v1/chat/completions
Or use the OpenAI-compatible proxy from Demeter-CodeBuilder, which wraps everything and exposes Bragi as a drop-in OpenAI endpoint.
How it works
The 1.5B Q3 backbone alone fails MBPP problems primarily by mis-recalling rare formulas: octagonal number written as 3n^2 - 2n instead of n(3n - 2), divisibility-by-11 implemented as digit-sum-mod-11 (correct rule is alternating digit sum). These formulas are 50-byte Python expressions; storing them in 3-bit quantised weights is wasteful and error-prone.
The full Bragi system externalises rare formulas into a 15 KB symbolic library (engine_lib.py) and routes matched problems entirely around the LLM via a regex keyword router. The backbone runs only when no route matches.
Quantisation details
- Base model:
Qwen/Qwen2.5-Coder-1.5B-Instruct - Quantisation: Q3_K_M with imatrix calibration
- Imatrix corpus: MBPP-train Python code split (374 verified solutions, 67 KB; no test data)
- Tool:
llama.cppllama-imatrix+llama-quantizebuild b4c0549 - File size: 786 MB
Limitations
- Hand-engineered engine_lib covers the MBPP test distribution but not arbitrary code tasks.
- Regex router will not match prompts in other languages or with unusual phrasing.
- Real-world coding (multi-file refactoring, integration with existing codebases) is not measured by MBPP and not the design target.
See paper section 5.3 for full limitations.
Citation
@misc{chen2026bragillm,
author = {Chen, Ho Yiing},
title = {Bragi-LLM: An 805 MB Hybrid Code-Generation System Reaches 92\% MBPP via LLM-Symbolic Engine Routing},
year = {2026},
doi = {10.5281/zenodo.20557449},
url = {https://doi.org/10.5281/zenodo.20557449},
note = {Independent Researcher, Taiwan. ORCID 0009-0006-6816-9891.}
}
Author
Chen, Ho Yiing (norika), Independent Researcher, Taiwan. ORCID: 0009-0006-6816-9891
Correspondence: norika at charenix.com
License
MIT. See the Bragi-LLM repo for the full LICENSE file.
Acknowledgements
Developed using donated off-hours access to NVIDIA DGX Spark hardware. Implementation and draft writing assisted by Claude (Anthropic). The architectural direction (refuse to ship sub-target results, diagnose before optimising, externalise rather than memorise) is the author's.