Agora 9B (GGUF)

Agora is a 9B multilingual assistant for European languages with an opt-in thinking mode and tool calling. It is the GGUF build of birgermoell/oellm-9b-256k-stage5-w2-rl-step160: the OpenEuroLLM 9B base model after instruct, reasoning and tool-use SFT, DPO, a weight merge and two reinforcement learning windows. The name comes from the agora, the square where people in ancient Greek cities met to talk and reason.

Research preview: good for trying European-language chat, step-by-step reasoning and tool use locally; not evaluated for deployment. See the full model card for lineage, training data, licences and every evaluation.

Run it

Ollama (the chat template, thinking switch, tool format and sampling settings come from this repo):

ollama run hf.co/birgermoell/agora-9b-GGUF:Q6_K            # recommended
ollama run hf.co/birgermoell/agora-9b-GGUF:Q6_K --think    # thinking mode: reason first, then answer
ollama run hf.co/birgermoell/agora-9b-GGUF:Q4_K_M          # smaller

In the Ollama app, pick the model and use the thinking toggle. Ollama turns thinking on by default for models that support it; switch it off for short, direct answers.

llama.cpp:

llama-server -hf birgermoell/agora-9b-GGUF:Q6_K --jinja --chat-template-file chat_template.jinja \
  --reasoning-format deepseek -c 32768
# thinking mode per request: "chat_template_kwargs": {"enable_thinking": true}

Use the chat_template.jinja from this repo (the training template without its training-only generation markers).

Converted with llama.cpp (convert_hf_to_gguf.py, commit a7af149) from the bf16 checkpoint. The tokenizer was checked token by token against the training tokenizer on rendered chat prompts (direct and thinking mode, tools, Swedish, Greek, Polish and Finnish text): 7/7 identical. Two settings differ from llama.cpp's defaults for this tokenizer type and are set in the files: no <bos> token (training sequences start with <|im_start|>) and no leading-space prefix.

What it is good at

Scores are from the full-precision model, greedy decoding (details and all other numbers in the full model card).

Known weaknesses: it rarely says "I don't know" (check names, dates and places); multi-step arithmetic over documents and tables is weak; refusal calibration is low; reasoning traces are mostly in English even when the answer is in another language.

Prompt format

ChatML with Qwen3/Hermes-style tools: <|im_start|>{role}\n{content}<|im_end|>\n. Thinking mode opens <think>\n at the start of the assistant turn and the model closes it with </think> before the answer. Tools are listed in the system turn inside <tools>…</tools>; the model calls them as <tool_call>\n{"name": …, "arguments": {…}}\n</tool_call>, and results come back as <tool_response> blocks in a user turn.

Licence and credits

Apache-2.0, like the base model. Built on the OpenEuroLLM 9B base model; post-training data licences and credits are listed in the full model card. Compute: EuroHPC JU / CINECA Leonardo. Code: qwen35-posttrain, oellm-rlvr.