---

language:

- en

- ar

license: apache-2.0

library_name: llama.cpp

pipeline_tag: text-generation

base_model: Qwen/Qwen3.5-9B

base_model_relation: finetune

datasets:

- NousResearch/hermes-function-calling-v1

- lambda/hermes-agent-reasoning-traces

- Roman1111111/gpt5.5-terminal

tags:

- gguf

- qwen3.5

- q4_k_m

- function-calling

- tool-use

- agent

- arabic

- fine-tuned

---

<p align="center">

<img src="assets/qwen35-anwar-hermes-cover.jpg" alt="Qwen3.5-9B-Anwar-Hermes model overview by AnwarPy" width="720">

</p>

Qwen3.5-9B-Anwar-Hermes — GGUF Q4_K_M

!Qwen3.5-9B-Anwar-Hermes performance overview

A compact GGUF release derived from Qwen/Qwen3.5-9B, fine-tuned for tool selection, structured output, function calling, and multi-step agent workflows in English and Arabic.

Built for Hermes Agent

Qwen3.5-9B-Anwar-Hermes is the result of a month-long training, dataset-curation, refinement, and evaluation cycle focused on practical compatibility with Hermes Agent and its tool ecosystem.

The model was shaped around function calling, structured tool schemas, terminal workflows, multi-step chains, and agent-style instructions. The goal is to provide a local model that works naturally with Hermes workflows—not a generic chat model with tool use added as an afterthought. Compatibility was validated against the documented Hermes-oriented behavior and terminal gates; it is not a claim of perfect parity across every tool, runtime, or future Hermes release.

At a glance

| | |

|---|---|

| Format | GGUF · Q4_K_M · 5.24 GiB |

| Base model | Qwen3.5-9B |

| Development cycle | One month of training, curation, refinement, and evaluation |

| Target integration | Hermes Agent tools and multi-step workflows |

| Validated context | 8,192 tokens |

| Primary strengths | Tool use, JSON, function calling, agent workflows |

| Languages | English and Arabic |

| Validated runtimes | llama.cpp, LM Studio, Windows llama.cpp |

Performance snapshot

| Evaluation | Result |

|---|---:|

| tool-eval-bench v2.1.0 | 91/100 · 126/138 points |

| Tool Selection | 100% |

| Multi-Step Chains | 100% |

| Structured Output | 100% |

| Context & State | 90% |

| Frozen behavior suite | 12/16 · threshold 11/16 |

| Terminal contract suite | 7/7 |

The benchmark covered 69 sequential scenarios with 52 tool definitions at temperature 0.0 and seed 42. These are task-specific results, not a claim of broad production readiness. The complete reproducibility summary is available in evals/tool-eval-bench-summary.md.

Quick start

1. Download

hf download AnwarPy/qwen3.5-9b-anwar-hermes-GGUF \
  qwen3.5-9b-anwar-hermes-Q4_K_M.gguf \
  --local-dir .

2. Start llama.cpp

llama-server \
  -m qwen3.5-9b-anwar-hermes-Q4_K_M.gguf \
  --alias qwen3.5-9b-anwar-hermes \
  -ngl 99 -c 8192 -np 1 \
  --host 127.0.0.1 --port 11442 \
  --jinja --reasoning off --reasoning-budget 0

3. Verify the API

curl http://127.0.0.1:11442/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "qwen3.5-9b-anwar-hermes",
    "messages": [{"role": "user", "content": "Reply with exactly: MODEL_OK"}],
    "temperature": 0
  }'

Expected assistant content:

MODEL_OK

The exact command above was validated against the published GGUF with llama.cpp commit e8bcb772d749a581d717b79ac8b9cd57301d9651. The same artifact passed load and inference canaries with official Windows llama.cpp b10078 (f04801018) and LM Studio engine 2.25.2. Runtime flags may vary in other versions.

Recommended use

• Function calling and tool selection.
• Strict or structured JSON generation.
• Multi-step local agent workflows.
• English and Arabic assistant tasks.
• Local OpenAI-compatible inference through llama.cpp-based runtimes.

This release is text/tool-use oriented. Vision capability inherited from the base model was not qualified for this GGUF release and should not be assumed.

Files and verification

| File | Purpose |

|---|---|

| qwen3.5-9b-anwar-hermes-Q4_K_M.gguf | Q4_K_M release · 5,629,108,704 bytes |

| SHA256SUMS | Release checksums |

| PROVENANCE.json | Machine-readable model, dataset, and compatibility lineage |

| data/terminal-curriculum-build-v2.json | Terminal curriculum build report |

| evals/tool-eval-bench-summary.md | Reproducible benchmark summary |

Release GGUF SHA-256:

cf0cc3541a75739861c4208b47a7f163c8273f0e56b4132b9bebf2a357a2c7c3

License

Apache License 2.0. See LICENSE. Users are also responsible for reviewing the base-model and dataset terms.

Acknowledgements

Built from Qwen3.5 by the Qwen team and trained with pinned data from Nous Research, the Hermes Agent reasoning-trace dataset, and the audited derived terminal curriculum sourced from Roman1111111/gpt5.5-terminal. Quantization and inference use the GGUF and llama.cpp ecosystem.