MiniMind-3 Mermaid Run 2 GGUF
GGUF files of the epoch 3 checkpoint (ep3/) of Lucasgitguy-co/minimind-mermaid-run2, a 64M-parameter Qwen3-architecture model tuned on Mermaid diagrams.
| File | Type | Size (GB) |
|---|---|---|
minimind-mermaid-run2-F16.gguf |
F16, same output as the original | 0.128 |
minimind-mermaid-run2-Q8_0.gguf |
Q8_0 | 0.068 |
minimind-mermaid-run2-Q4_K_M.gguf |
Q4_K_M | 0.042 |
Use
The chat template (minimind-3, ChatML) is embedded, so chat mode works:
hf download Lucasgitguy-co/minimind-mermaid-run2-GGUF minimind-mermaid-run2-Q8_0.gguf --local-dir .
llama-cli -m minimind-mermaid-run2-Q8_0.gguf --jinja -p "Write Mermaid code for a flowchart: Start, Process, End."
This is a very small model. Its diagrams are often wrong or off-topic: in the tests below it drew pie charts for flowchart and sequence prompts. Use it for experiments.
How these were made and tested
- Converted from
ep3/with llama.cppconvert_hf_to_gguf.py(checkout1b6935b). The token ids and chat template come from open PR #1 of the source repo. - The converter doesn't know this tokenizer's hash, so a wrapper (
convert_with_pre.py) maps it togpt-2. That matches itsByteLevelpre-tokenizer with the GPT-2 regex. - Tested 2026-11-16 with llama.cpp 1.0.0 (build 12359):
- Token ids of 3 chat-formatted prompts match transformers.
- Greedy output of the F16 file is identical to transformers (float32) on those 3 prompts, 60 tokens each.
- Q8_0 and Q4_K_M were spot-checked on one prompt. The first lines match F16, then they drift.
Licence
Apache-2.0. Based on jingyaogong/minimind-3 (Apache-2.0). See NOTICE and licenses/.