---

license: apache-2.0

base_model:

- openbmb/MiniCPM5-2B

base_model_relation: finetune

library_name: gguf

pipeline_tag: text-generation

tags: [gguf, llama-cpp, ollama, minicpm5, elffuss, coding-agent, tool-calling, q8_0, on-device]

language: [es, en]

---

!banner

<div style="border:2px solid currentColor; font-family:ui-monospace,'SF Mono','Cascadia Mono',Consolas,monospace;">

<div style="border-bottom:1px solid currentColor; padding:6px 12px; font-size:11px; letter-spacing:3px; text-transform:uppercase; opacity:0.7; text-align:center;">KIKOCIS // ELFFUSS CODE // A NEW VERSION EVERY DAY</div>

<div style="padding:14px; display:flex; flex-wrap:wrap; align-items:center; justify-content:center; gap:18px;">

<pre style="margin:0; font-size:9px; line-height:1.15;">

you ──▶ ``tool code.search ``

[resultado] lib/x.js:42

──▶ ``tool code.edit ``

✔ fixed, verified

</pre>

<div style="text-align:center;">

<div style="font-size:23px; font-weight:800; letter-spacing:1px;">MiniCPM5-2B · Elffuss Code</div>

<div style="font-size:12.5px; letter-spacing:1px; opacity:0.8; margin-top:5px;"><span style="white-space:nowrap;">2B · Q8_0-XL</span> · <span style="white-space:nowrap;">v7 · 2026-10-17</span> · <span style="white-space:nowrap;">bugs fixed 4 %</span></div>

</div>

</div>

</div>

MiniCPM5-2B — Elffuss Code edition (GGUF)

> OpenBMB's MiniCPM5-2B, trained a little more every day to work as the coding agent of Elffuss Code — the IDE that runs in your browser: read the project, find the bug, edit the exact lines, check the fix, answer citing real files. A new version is published every day, and every version is measured on the same held-out tasks, so the whole evolution is below.

🌐 Try it in Elffuss Code → elffuss.com

📈 Evolution

| Version | Date | Bugs fixed | Grounded answers | Valid tool protocol | Finished | Avg. steps | KLD vs original |

|---|---|---:|---:|---:|---:|---:|---:|

| v0 (original) | — | 12 % | 92 % | 98 % | 77 % | 6.1 | 0.0000 |

| v1 | 2026-10-11 | 0 % | 0 % | 85 % | 100 % | 1 | 8.7960 |

| v2 | 2026-10-11 | 4 % | 33 % | 85 % | 17 % | 10.38 | 0.0494 |

| v3 | 2026-10-13 | 0 % | 25 % | 92 % | 21 % | 9.75 | 0.0204 |

| v4 | 2026-10-14 | 4 % | 58 % | 96 % | 35 % | 8.94 | 0.0211 |

| v5 | 2026-10-15 | 0 % | 42 % | 88 % | 62 % | 7.27 | 0.0214 |

| v6 | 2026-10-16 | 8 % | 33 % | 90 % | 42 % | 8.38 | 0.0224 |

| v7 | 2026-10-17 | 4 % | 25 % | 85 % | 35 % | 8.94 | 0.0209 |

!evolution

• Bugs fixed — a one-token bug is planted in a real function of a repo the model never trained on; it counts only if the file ends byte-identical to the original.
• Grounded answers — explaining or locating code, the answer cites a file the tools actually returned.
• Valid tool protocol — every tool call is one Elffuss Code's own parser accepts.
• KLD vs original — drift of the next-token distribution from the original BF16 model on general text (wikitext-2): how much general behaviour the specialisation costs.

Held-out set: 24 tasks on 3 public repos that are never used for training, each run twice, sampled at temperature 1.0 / top_p 0.95 in Elffuss Code's own agent loop (its system prompt, tools and parser, unchanged).

🚀 Run it

llama-server -m MiniCPM5-2B-Elffuss-Code-Q8_0-XL.gguf -c 32768 --jinja --temp 1.0 --top-p 0.95 \
  --chat-template-kwargs '{"enable_thinking": false}'
ollama run hf.co/KikoCis/MiniCPM5-2B-Elffuss-Code-GGUF:Q8_0-XL

Built for Elffuss Code's tool protocol (``` `tool {"tool": "code.read", …} ` ``` blocks). In Elffuss Code, point the API provider at your llama-server.

📦 File

| File | Quant | Size |

|---|---|---:|

| MiniCPM5-2B-Elffuss-Code-Q8_0-XL.gguf | Q8_0 blocks, BF16 embeddings + output head | 3.18 GB |

Every daily version is a git tag on this repo (v<date>), so any day can be downloaded again.

📚 Credit & license

Base model © OpenBMB — openbmb/MiniCPM5-2B. Elffuss Code specialisation, daily builds and measurements: KikoCis — elffuss.com. Apache-2.0, same as upstream. The faithful (unmodified) quantization is at KikoCis/MiniCPM5-2B-GGUF.