Instructions to use tiny-random/qwen3.8-moe with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use tiny-random/qwen3.8-moe with Transformers:

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

pipe = pipeline("text-generation", model="tiny-random/qwen3.8-moe")
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("tiny-random/qwen3.8-moe")
model = AutoModelForCausalLM.from_pretrained("tiny-random/qwen3.8-moe", 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 tiny-random/qwen3.8-moe with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "tiny-random/qwen3.8-moe"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "tiny-random/qwen3.8-moe",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/tiny-random/qwen3.8-moe

SGLang

How to use tiny-random/qwen3.8-moe 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 "tiny-random/qwen3.8-moe" \
    --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": "tiny-random/qwen3.8-moe",
		"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 "tiny-random/qwen3.8-moe" \
        --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": "tiny-random/qwen3.8-moe",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use tiny-random/qwen3.8-moe with Docker Model Runner:
```
docker model run hf.co/tiny-random/qwen3.8-moe
```

This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from Qwen/Qwen3.8-2.4T-A95B.

Structural comparison with Qwen/Qwen3.8-2.4T-A95B:

Feature	Source	Tiny
Input modalities	Text	Same
Vocabulary	248,320	Same; special-token IDs preserved
Backbone layers	92	8
Hidden size	8,192	16
Attention Q/KV heads; head dim	64/4; 256	32/2; 128
Attention cycle	3 linear + 1 full	Same; two complete cycles
Linear-attention key/value heads	16/128	2/16
Linear-attention key/value head dim	128/128	Same
Routed experts / top-k	512 / 10	32 / 10
Shared experts	1	Same
Expert/shared intermediate size	2,048 / 2,048	256 / 256
MTP	1 full-attention layer; shared embeddings	Same
Checkpoint format	BF16	Same

File path	Size
model.safetensors	25.9MB

Example usage:

SGLang (✅ verified)
- Runs with TP=2, the embedded NEXTN/MTP layer, and CUDA Graph.
- Verified for text requests on two NVIDIA H20 GPUs; see Test environment for dependency versions.

model_id=tiny-random/qwen3.8-moe
sglang serve \
  --model-path $model_id \
  --tp-size 2 \
  --fp8-gemm-backend triton \
  --moe-runner-backend triton \
  --speculative-algorithm NEXTN \
  --speculative-num-steps 3 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 4 \
  --context-length 4096 \
  --max-total-tokens 4096 \
  --chunked-prefill-size 512 \
  --mem-fraction-static 0.4 \
  --cuda-graph-config '{"decode": {"bs": [1, 2, 4, 8]}}' \
  --disable-custom-all-reduce

Transformers (✅ verified)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "tiny-random/qwen3.8-moe"
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="cuda", attn_implementation="sdpa")
tokenizer = AutoTokenizer.from_pretrained(model_id)
inputs = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Hello"}], tokenize=True,
    add_generation_prompt=True, return_dict=True, return_tensors="pt",
).to(model.device)
output = model.generate(**inputs, do_sample=False, max_new_tokens=10)
print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Codes to create this repo (✅ verified):

Click to expand

import json
from copy import deepcopy
from pathlib import Path
from shutil import copyfile

import torch
from huggingface_hub import hf_hub_download, list_repo_files
from safetensors.torch import save_file
from transformers import AutoConfig, AutoModelForCausalLM, set_seed

source_model_id = "Qwen/Qwen3.8-2.4T-A95B"
save_folder = "/tmp/tiny-random/qwen38-moe"
source_revision = '207bd685a7e3696cfaff12ded7c6a7ea0f88c996'
set_seed(42)
folder = Path(save_folder)
folder.mkdir(parents=True, exist_ok=True)
raw = json.loads(Path(hf_hub_download(source_model_id, "config.json", revision=source_revision)).read_text())
raw.update(
    hidden_size=16, head_dim=128, num_hidden_layers=8,
    layer_types=(["linear_attention"] * 3 + ["full_attention"]) * 2,
    num_attention_heads=32, num_key_value_heads=2,
    linear_num_key_heads=2, linear_num_value_heads=16,
    linear_key_head_dim=128, linear_value_head_dim=128,
    num_experts=32, moe_intermediate_size=256, shared_expert_intermediate_size=256,
)
(folder / "config.json").write_text(json.dumps(raw, indent=2))
config = AutoConfig.from_pretrained(folder)
model = AutoModelForCausalLM.from_config(config, dtype=torch.bfloat16)
backbone = model.model
text = backbone.config
def norm():
    return type(backbone.norm)(text.hidden_size, text.rms_norm_eps).to(torch.bfloat16)
model.mtp = torch.nn.ModuleDict({
    "pre_fc_norm_embedding": norm(),
    "pre_fc_norm_hidden": norm(),
    "fc": torch.nn.Linear(2 * text.hidden_size, text.hidden_size, bias=False, dtype=torch.bfloat16),
    "layers": torch.nn.ModuleList([deepcopy(backbone.layers[-1])]),
    "norm": norm(),
})
with torch.no_grad():
    for name, param in sorted(model.named_parameters()):
        if name.endswith("A_log"):
            param.zero_()
        elif name.endswith("dt_bias"):
            param.fill_(1)
        elif "norm" in name and name.endswith("weight"):
            param.fill_(1 if "visual" in name or ".linear_attn.norm." in name else 0)
        elif name.endswith("bias"):
            param.zero_()
        else:
            torch.nn.init.normal_(param, mean=0, std=0.02)
state = {name: value.detach().cpu().contiguous() for name, value in model.state_dict().items()}
save_file(state, folder / "model.safetensors", metadata={"format": "pt"})
# Preserve tokenizer, processor and generation assets exactly as published.
assets = {
    "tokenizer.json", "tokenizer_config.json", "special_tokens_map.json",
    "vocab.json", "merges.txt", "added_tokens.json", "chat_template.jinja",
    "chat_template.json", "preprocessor_config.json", "processor_config.json",
    "video_preprocessor_config.json", "generation_config.json",
}
for name in sorted(set(list_repo_files(source_model_id, revision=source_revision)) & assets):
    copyfile(hf_hub_download(source_model_id, name, revision=source_revision), folder / name)

Printing the model:

Click to expand

Qwen3_5MoeForCausalLM(
  (model): Qwen3_5MoeTextModel(
    (embed_tokens): Embedding(248320, 16)
    (layers): ModuleList(
      (0-2): 3 x Qwen3_5MoeDecoderLayer(
        (linear_attn): Qwen3_5MoeGatedDeltaNet(
          (conv1d): Conv1d(2560, 2560, kernel_size=(4,), stride=(1,), padding=(3,), groups=2560, bias=False)
          (norm): Qwen3_5MoeRMSNormGated()
          (out_proj): Linear(in_features=2048, out_features=16, bias=False)
          (in_proj_qkv): Linear(in_features=16, out_features=2560, bias=False)
          (in_proj_z): Linear(in_features=16, out_features=2048, bias=False)
          (in_proj_b): Linear(in_features=16, out_features=16, bias=False)
          (in_proj_a): Linear(in_features=16, out_features=16, bias=False)
        )
        (mlp): Qwen3_5MoeSparseMoeBlock(
          (gate): Qwen3_5MoeTopKRouter()
          (experts): Qwen3_5MoeExperts(
            (act_fn): SiLUActivation()
          )
          (shared_expert): Qwen3_5MoeMLP(
            (gate_proj): Linear(in_features=16, out_features=256, bias=False)
            (up_proj): Linear(in_features=16, out_features=256, bias=False)
            (down_proj): Linear(in_features=256, out_features=16, bias=False)
            (act_fn): SiLUActivation()
          )
          (shared_expert_gate): Linear(in_features=16, out_features=1, bias=False)
        )
        (input_layernorm): Qwen3_5MoeRMSNorm((16,), eps=1e-06)
        (post_attention_layernorm): Qwen3_5MoeRMSNorm((16,), eps=1e-06)
      )
      (3): Qwen3_5MoeDecoderLayer(
        (self_attn): Qwen3_5MoeAttention(
          (q_proj): Linear(in_features=16, out_features=8192, bias=False)
          (k_proj): Linear(in_features=16, out_features=256, bias=False)
          (v_proj): Linear(in_features=16, out_features=256, bias=False)
          (o_proj): Linear(in_features=4096, out_features=16, bias=False)
          (q_norm): Qwen3_5MoeRMSNorm((128,), eps=1e-06)
          (k_norm): Qwen3_5MoeRMSNorm((128,), eps=1e-06)
        )
        (mlp): Qwen3_5MoeSparseMoeBlock(
          (gate): Qwen3_5MoeTopKRouter()
          (experts): Qwen3_5MoeExperts(
            (act_fn): SiLUActivation()
          )
          (shared_expert): Qwen3_5MoeMLP(
            (gate_proj): Linear(in_features=16, out_features=256, bias=False)
            (up_proj): Linear(in_features=16, out_features=256, bias=False)
            (down_proj): Linear(in_features=256, out_features=16, bias=False)
            (act_fn): SiLUActivation()
          )
          (shared_expert_gate): Linear(in_features=16, out_features=1, bias=False)
        )
        (input_layernorm): Qwen3_5MoeRMSNorm((16,), eps=1e-06)
        (post_attention_layernorm): Qwen3_5MoeRMSNorm((16,), eps=1e-06)
      )
      (4-6): 3 x Qwen3_5MoeDecoderLayer(
        (linear_attn): Qwen3_5MoeGatedDeltaNet(
          (conv1d): Conv1d(2560, 2560, kernel_size=(4,), stride=(1,), padding=(3,), groups=2560, bias=False)
          (norm): Qwen3_5MoeRMSNormGated()
          (out_proj): Linear(in_features=2048, out_features=16, bias=False)
          (in_proj_qkv): Linear(in_features=16, out_features=2560, bias=False)
          (in_proj_z): Linear(in_features=16, out_features=2048, bias=False)
          (in_proj_b): Linear(in_features=16, out_features=16, bias=False)
          (in_proj_a): Linear(in_features=16, out_features=16, bias=False)
        )
        (mlp): Qwen3_5MoeSparseMoeBlock(
          (gate): Qwen3_5MoeTopKRouter()
          (experts): Qwen3_5MoeExperts(
            (act_fn): SiLUActivation()
          )
          (shared_expert): Qwen3_5MoeMLP(
            (gate_proj): Linear(in_features=16, out_features=256, bias=False)
            (up_proj): Linear(in_features=16, out_features=256, bias=False)
            (down_proj): Linear(in_features=256, out_features=16, bias=False)
            (act_fn): SiLUActivation()
          )
          (shared_expert_gate): Linear(in_features=16, out_features=1, bias=False)
        )
        (input_layernorm): Qwen3_5MoeRMSNorm((16,), eps=1e-06)
        (post_attention_layernorm): Qwen3_5MoeRMSNorm((16,), eps=1e-06)
      )
      (7): Qwen3_5MoeDecoderLayer(
        (self_attn): Qwen3_5MoeAttention(
          (q_proj): Linear(in_features=16, out_features=8192, bias=False)
          (k_proj): Linear(in_features=16, out_features=256, bias=False)
          (v_proj): Linear(in_features=16, out_features=256, bias=False)
          (o_proj): Linear(in_features=4096, out_features=16, bias=False)
          (q_norm): Qwen3_5MoeRMSNorm((128,), eps=1e-06)
          (k_norm): Qwen3_5MoeRMSNorm((128,), eps=1e-06)
        )
        (mlp): Qwen3_5MoeSparseMoeBlock(
          (gate): Qwen3_5MoeTopKRouter()
          (experts): Qwen3_5MoeExperts(
            (act_fn): SiLUActivation()
          )
          (shared_expert): Qwen3_5MoeMLP(
            (gate_proj): Linear(in_features=16, out_features=256, bias=False)
            (up_proj): Linear(in_features=16, out_features=256, bias=False)
            (down_proj): Linear(in_features=256, out_features=16, bias=False)
            (act_fn): SiLUActivation()
          )
          (shared_expert_gate): Linear(in_features=16, out_features=1, bias=False)
        )
        (input_layernorm): Qwen3_5MoeRMSNorm((16,), eps=1e-06)
        (post_attention_layernorm): Qwen3_5MoeRMSNorm((16,), eps=1e-06)
      )
    )
    (norm): Qwen3_5MoeRMSNorm((16,), eps=1e-06)
    (rotary_emb): Qwen3_5MoeTextRotaryEmbedding()
  )
  (lm_head): Linear(in_features=16, out_features=248320, bias=False)
  (mtp): ModuleDict(
    (pre_fc_norm_embedding): Qwen3_5MoeRMSNorm((16,), eps=1e-06)
    (pre_fc_norm_hidden): Qwen3_5MoeRMSNorm((16,), eps=1e-06)
    (fc): Linear(in_features=32, out_features=16, bias=False)
    (layers): ModuleList(
      (0): Qwen3_5MoeDecoderLayer(
        (self_attn): Qwen3_5MoeAttention(
          (q_proj): Linear(in_features=16, out_features=8192, bias=False)
          (k_proj): Linear(in_features=16, out_features=256, bias=False)
          (v_proj): Linear(in_features=16, out_features=256, bias=False)
          (o_proj): Linear(in_features=4096, out_features=16, bias=False)
          (q_norm): Qwen3_5MoeRMSNorm((128,), eps=1e-06)
          (k_norm): Qwen3_5MoeRMSNorm((128,), eps=1e-06)
        )
        (mlp): Qwen3_5MoeSparseMoeBlock(
          (gate): Qwen3_5MoeTopKRouter()
          (experts): Qwen3_5MoeExperts(
            (act_fn): SiLUActivation()
          )
          (shared_expert): Qwen3_5MoeMLP(
            (gate_proj): Linear(in_features=16, out_features=256, bias=False)
            (up_proj): Linear(in_features=16, out_features=256, bias=False)
            (down_proj): Linear(in_features=256, out_features=16, bias=False)
            (act_fn): SiLUActivation()
          )
          (shared_expert_gate): Linear(in_features=16, out_features=1, bias=False)
        )
        (input_layernorm): Qwen3_5MoeRMSNorm((16,), eps=1e-06)
        (post_attention_layernorm): Qwen3_5MoeRMSNorm((16,), eps=1e-06)
      )
    )
    (norm): Qwen3_5MoeRMSNorm((16,), eps=1e-06)
  )
)

Test environment:

accelerate: 1.15.0
flashinfer-python: 0.7.2
huggingface_hub: 1.33.0
kernels: 0.17.0
safetensors: 0.8.0
sglang: 0.5.23.dev384+g9ab74b851
torch: 2.14.1
transformers: 5.19.0

Change log:

2026-10-23: Initial version.
2026-10-23: Reduced all repo files below 50 MB; GPU-validated TP=2, MTP and CUDA Graph.
2026-10-23: Creation examples use only public dependencies; source assets are copied unchanged.

This repo is generated by GPT-6.1-Sol