---

license: mit

base_model: deepreinforce-ai/Ornith-1.0-9B

tags:

• gguf
• ollama
• llama.cpp
• agentic-coding
• chat-template-fix

language:

• en

---

!banner

Ornith-1.0-9B — chat-template fixed (Q8_0 GGUF)

> This is NOT my model. All weights and training are by DeepReinforce (deepreinforce-ai/Ornith-1.0-9B). This is an independent repack that fixes one packaging bug and adds an honest third-party evaluation. No weights were modified.

The bug this fixes 🐛

The official Ornith GGUFs (and the base Qwen3.5-9B GGUFs) ship without an embedded tokenizer.chat_template. As a result, Ollama and llama.cpp fall back to a raw passthrough template ({{ .Prompt }}), the model never sees its ChatML format, and it degenerates into repetition loops — input tokens freeze, the same phrase repeats, the agent never acts.

The fix: this GGUF embeds DeepReinforce's official chat_template.jinja into the file metadata, so it works out-of-the-box in Ollama (auto-detected) and llama.cpp (--jinja). A bug-report note is also embedded in general.description.

📣 DeepReinforce: please add tokenizer.chat_template to your official GGUF exports — it's a one-line metadata fix that makes the GGUFs usable.

Independent evaluation (honest, small probe)

6-instance SWE-bench (django) probe, Claude-Code agentic harness, temp 1.0 / top_p 0.95 (DeepReinforce's recommended sampling — see below), same template for all:

| model | SWE | note |

|---|---|---|

| base Qwen3.5-9B | 1/6 | baseline |

| Ornith-1.0-9B | 2/6 | real fine-tune, beats base |

| NRS_QWEN_MYTHOS_1M ("100x reasoning" hype) | 0/6 | hype, worse than base |

And the two config bugs that matter for ALL reasoning GGUFs:

| Ornith-9B, only the change shown | result |

|---|---|

| temp 0.1, raw template | 0/6 (repetition loops) |

| temp 1.0 + this fixed template | 2/6 (healthy) |

> ⚠️ These are relative numbers on a tiny probe — NOT comparable 1:1 to DeepReinforce's official 69.4 SWE-bench Verified (measured on the OpenHands harness with their <function=> tool format, temp 1.0, top_p 0.95). Use the official numbers for leaderboard comparison. The point here is: the model works once configured right.

Full write-up: see FINDINGS.md.

Usage

Ollama:

ollama run hf.co/KikoCis/Ornith-1.0-9B-Ollama-fixed-GGUF
# or: ollama create ornith -f Modelfile && ollama run ornith

llama.cpp:

llama-server -m ornith-1.0-9b-Q8_0-fixed.gguf --jinja -c 32768

Recommended sampling (DeepReinforce official): temperature 1.0, top_p 0.95. Low temperature (e.g. 0.1) causes repetition loops on this model — use ~1.0.

Credit & license

• Model, weights, training, and the chat_template.jinja: © DeepReinforce — Ornith-1.0, GrandCode paper, deep-reinforce.com.
• This repack + independent evaluation: KikoCis. MIT (same as upstream). No weights modified — only metadata (chat_template, general.description) added.