Instructions to use eadx/Qwythos-9B-Claude-Mythos-5-1M with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use eadx/Qwythos-9B-Claude-Mythos-5-1M with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="eadx/Qwythos-9B-Claude-Mythos-5-1M")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("eadx/Qwythos-9B-Claude-Mythos-5-1M")
model = AutoModelForMultimodalLM.from_pretrained("eadx/Qwythos-9B-Claude-Mythos-5-1M")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use eadx/Qwythos-9B-Claude-Mythos-5-1M with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "eadx/Qwythos-9B-Claude-Mythos-5-1M"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "eadx/Qwythos-9B-Claude-Mythos-5-1M",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/eadx/Qwythos-9B-Claude-Mythos-5-1M

SGLang

How to use eadx/Qwythos-9B-Claude-Mythos-5-1M with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "eadx/Qwythos-9B-Claude-Mythos-5-1M" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "eadx/Qwythos-9B-Claude-Mythos-5-1M",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "eadx/Qwythos-9B-Claude-Mythos-5-1M" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "eadx/Qwythos-9B-Claude-Mythos-5-1M",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use eadx/Qwythos-9B-Claude-Mythos-5-1M with Docker Model Runner:
```
docker model run hf.co/eadx/Qwythos-9B-Claude-Mythos-5-1M
```

Qwythos-9B-Claude-Mythos-5-1M / evals /lm_eval_results.md

eadx

Duplicate from empero-ai/Qwythos-9B-Claude-Mythos-5-1M

a889ddd 9 days ago

preview code

Raw

History Blame Contribute Delete

3.15 kB

	# Qwythos-9B vs. base Qwen3.5-9B — lm-evaluation-harness

	Generative reasoning + broad-knowledge comparison under identical evaluation conditions (same harness, same backend, same prompts, same sampling). Run with the official [`lm-evaluation-harness`](https://github.com/EleutherAI/lm-evaluation-harness), HF backend, batch_size=auto, `--apply_chat_template`, Qwen3.5 sampling (`max_gen_toks=8192, temperature=0.6, top_p=0.95, top_k=20`), `--limit 100`.

	## Headline results

	\| Task \| Metric \| Base Qwen3.5-9B \| Qwythos-9B \| Δ \|
	\|---\|---\|---:\|---:\|---:\|
	\| gsm8k \| exact_match (flexible-extract) \| 0.670 \| 0.860 \| +0.190 \|
	\| gsm8k \| exact_match (strict-match) \| 0.510 \| 0.810 \| +0.300 \|
	\| mmlu \| acc \| 0.232 \| 0.575 \| +0.343 \|
	\| arc_challenge \| acc \| 0.470 \| 0.490 \| +0.020 \|
	\| arc_challenge \| acc_norm \| 0.400 \| 0.410 \| +0.010 \|
	\| gpqa_diamond_cot_zeroshot \| exact_match (flexible) \| 0.630 \| 0.580 \| −0.050 \|
	\| gpqa_diamond_cot_zeroshot \| exact_match (strict) \| 0.050 \| 0.010 \| −0.040 \|

	See [`assets/qwythos_eval_chart.svg`](../assets/qwythos_eval_chart.svg) for a visualization.

	## MMLU — domain breakdown (Qwythos, mean over 57 subjects)

	\| Domain \| Mean accuracy \| Subjects \|
	\|---\|---:\|---:\|
	\| Social sciences \| 0.667 \| 12 \|
	\| Other (business / med-adjacent / applied) \| 0.629 \| 8 \|
	\| STEM \| 0.544 \| 18 \|
	\| Medical \| 0.525 \| 6 \|
	\| Humanities \| 0.521 \| 13 \|

	Aggregate MMLU 0.575 — a +34.3-point lift over base under matched evaluation.

	## Reading these numbers honestly

	- *The wins are large and they are real under identical evaluation conditions.* Both models were evaluated with the exact same harness, prompts, sampling, and `--apply_chat_template` setting. Differences reflect differences in the model, not in the setup.
	- gsm8k +30 pts strict is the cleanest signal — same harness, same sampling, same extraction. The model is meaningfully stronger at math reasoning.
	- MMLU +34.3 is the broad-knowledge headline. Absolute MMLU numbers for Qwen3.5-9B vary considerably across evaluation pipelines (harness choice, few-shot count, chat-template handling, sampling); the matched-condition delta is what's meaningful here.
	- gpqa flexible-extract −5 pts is the one small regression — graduate-physics reasoning narrowed slightly. The strict numbers (0.05 / 0.01) for both models are degenerate — both fail the regex extractor; the flex score is the meaningful one.

	## Reproducing

	```bash
	git clone https://github.com/EleutherAI/lm-evaluation-harness
	cd lm-evaluation-harness && pip install -e ".[math,ifeval]"

	lm_eval --model hf \
	--model_args pretrained=empero-ai/Qwythos-9B-Claude-Mythos-5-1M,dtype=bfloat16,trust_remote_code=True,max_length=16384 \
	--tasks gsm8k,minerva_math,gpqa_diamond_cot_zeroshot,mmlu,arc_challenge \
	--apply_chat_template \
	--gen_kwargs "max_gen_toks=8192,temperature=0.6,top_p=0.95,top_k=20,do_sample=true" \
	--batch_size auto --limit 100 \
	--output_path qwythos_eval
	```

	GPQA requires HF dataset access (gated); request it once at [Idavidrein/gpqa](https://huggingface.co/datasets/Idavidrein/gpqa).

	# Qwythos-9B vs. base Qwen3.5-9B — lm-evaluation-harness

	Generative reasoning + broad-knowledge comparison under identical evaluation conditions (same harness, same backend, same prompts, same sampling). Run with the official [`lm-evaluation-harness`](https://github.com/EleutherAI/lm-evaluation-harness), HF backend, batch_size=auto, `--apply_chat_template`, Qwen3.5 sampling (`max_gen_toks=8192, temperature=0.6, top_p=0.95, top_k=20`), `--limit 100`.

	## Headline results

	\| Task \| Metric \| Base Qwen3.5-9B \| Qwythos-9B \| Δ \|
	\|---\|---\|---:\|---:\|---:\|
	\| gsm8k \| exact_match (flexible-extract) \| 0.670 \| 0.860 \| +0.190 \|
	\| gsm8k \| exact_match (strict-match) \| 0.510 \| 0.810 \| +0.300 \|
	\| mmlu \| acc \| 0.232 \| 0.575 \| +0.343 \|
	\| arc_challenge \| acc \| 0.470 \| 0.490 \| +0.020 \|
	\| arc_challenge \| acc_norm \| 0.400 \| 0.410 \| +0.010 \|
	\| gpqa_diamond_cot_zeroshot \| exact_match (flexible) \| 0.630 \| 0.580 \| −0.050 \|
	\| gpqa_diamond_cot_zeroshot \| exact_match (strict) \| 0.050 \| 0.010 \| −0.040 \|

	See [`assets/qwythos_eval_chart.svg`](../assets/qwythos_eval_chart.svg) for a visualization.

	## MMLU — domain breakdown (Qwythos, mean over 57 subjects)

	\| Domain \| Mean accuracy \| Subjects \|
	\|---\|---:\|---:\|
	\| Social sciences \| 0.667 \| 12 \|
	\| Other (business / med-adjacent / applied) \| 0.629 \| 8 \|
	\| STEM \| 0.544 \| 18 \|
	\| Medical \| 0.525 \| 6 \|
	\| Humanities \| 0.521 \| 13 \|

	Aggregate MMLU 0.575 — a +34.3-point lift over base under matched evaluation.

	## Reading these numbers honestly

	- *The wins are large and they are real under identical evaluation conditions.* Both models were evaluated with the exact same harness, prompts, sampling, and `--apply_chat_template` setting. Differences reflect differences in the model, not in the setup.
	- gsm8k +30 pts strict is the cleanest signal — same harness, same sampling, same extraction. The model is meaningfully stronger at math reasoning.
	- MMLU +34.3 is the broad-knowledge headline. Absolute MMLU numbers for Qwen3.5-9B vary considerably across evaluation pipelines (harness choice, few-shot count, chat-template handling, sampling); the matched-condition delta is what's meaningful here.
	- gpqa flexible-extract −5 pts is the one small regression — graduate-physics reasoning narrowed slightly. The strict numbers (0.05 / 0.01) for both models are degenerate — both fail the regex extractor; the flex score is the meaningful one.

	## Reproducing

	```bash
	git clone https://github.com/EleutherAI/lm-evaluation-harness
	cd lm-evaluation-harness && pip install -e ".[math,ifeval]"

	lm_eval --model hf \
	--model_args pretrained=empero-ai/Qwythos-9B-Claude-Mythos-5-1M,dtype=bfloat16,trust_remote_code=True,max_length=16384 \
	--tasks gsm8k,minerva_math,gpqa_diamond_cot_zeroshot,mmlu,arc_challenge \
	--apply_chat_template \
	--gen_kwargs "max_gen_toks=8192,temperature=0.6,top_p=0.95,top_k=20,do_sample=true" \
	--batch_size auto --limit 100 \
	--output_path qwythos_eval
	```

	GPQA requires HF dataset access (gated); request it once at [Idavidrein/gpqa](https://huggingface.co/datasets/Idavidrein/gpqa).