Instructions to use tn-mixt/soup10x1.5 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use tn-mixt/soup10x1.5 with Transformers:

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

pipe = pipeline("text-generation", model="tn-mixt/soup10x1.5", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("tn-mixt/soup10x1.5", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("tn-mixt/soup10x1.5", trust_remote_code=True, device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use tn-mixt/soup10x1.5 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "tn-mixt/soup10x1.5"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "tn-mixt/soup10x1.5",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/tn-mixt/soup10x1.5

SGLang

How to use tn-mixt/soup10x1.5 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 "tn-mixt/soup10x1.5" \
    --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": "tn-mixt/soup10x1.5",
		"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 "tn-mixt/soup10x1.5" \
        --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": "tn-mixt/soup10x1.5",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use tn-mixt/soup10x1.5 with Docker Model Runner:
```
docker model run hf.co/tn-mixt/soup10x1.5
```

Qwen3.5-MixT 3.5B Math (256k)

Private math-reasoning checkpoint for the tn-mixt Qwen3.5-MixT 3.5B model.

Architecture: Qwen3_5MixTForCausalLM, MixT tail13, 3.40B unique parameters, bf16
Context length: 262,144 tokens
Repository: private, members of tn-mixt

Evaluation

AIME24 / AIME25 / AIME26, 256k generation budget, 10 samples per problem (Avg@10):

Set	Accuracy
AIME24	91.0%
AIME25	89.3%
AIME26	89.0%
ALL	89.8%

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "tn-mixt/soup10x1.5"
tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    torch_dtype="bfloat16",
    device_map="auto",
)

The repository includes the custom modeling files (modeling_qwen3_5_mixt.py, mixt_projector.py, mixt_manifest.json); trust_remote_code=True is required. For vLLM serving, install the MixT plugin and point it at this directory.

Files

File	Description
`model.safetensors`	bf16 weights (~6.33 GiB)
`config.json`, `generation_config.json`	model config
`tokenizer.json`, `tokenizer_config.json`, `chat_template.jinja`	tokenizer
`modeling_qwen3_5_mixt.py`, `mixt_projector.py`, `mixt_manifest.json`	custom MixT code

Limitations

Evaluated on AIME only; results are from a single 90-problem run and have a standard error of roughly 1.2 points.
Math-only checkpoint; no RL/verifier training. Long generations can occasionally run to the 256k context limit on difficult problems.

Downloads last month: 10

Safetensors

Model size

3B params

Tensor type

BF16