Instructions to use ManniX-ITA/Qwen3.6-27B-A3B-CoderX with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ManniX-ITA/Qwen3.6-27B-A3B-CoderX with Transformers:

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

pipe = pipeline("text-generation", model="ManniX-ITA/Qwen3.6-27B-A3B-CoderX")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("ManniX-ITA/Qwen3.6-27B-A3B-CoderX")
model = AutoModelForCausalLM.from_pretrained("ManniX-ITA/Qwen3.6-27B-A3B-CoderX", 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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use ManniX-ITA/Qwen3.6-27B-A3B-CoderX with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ManniX-ITA/Qwen3.6-27B-A3B-CoderX"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ManniX-ITA/Qwen3.6-27B-A3B-CoderX",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ManniX-ITA/Qwen3.6-27B-A3B-CoderX

SGLang

How to use ManniX-ITA/Qwen3.6-27B-A3B-CoderX 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 "ManniX-ITA/Qwen3.6-27B-A3B-CoderX" \
    --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": "ManniX-ITA/Qwen3.6-27B-A3B-CoderX",
		"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 "ManniX-ITA/Qwen3.6-27B-A3B-CoderX" \
        --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": "ManniX-ITA/Qwen3.6-27B-A3B-CoderX",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ManniX-ITA/Qwen3.6-27B-A3B-CoderX with Docker Model Runner:
```
docker model run hf.co/ManniX-ITA/Qwen3.6-27B-A3B-CoderX
```

Qwen3.6-27B-A3B-CoderX

BF16 weights for CoderX — a long-horizon code prune of Qwen/Qwen3.6-35B-A3B: 256 experts per layer reduced to 184, ~35B → ~27B, still A3B active.

Same expert budget as the sibling Qwen3.6-27B-A3B-Coder, but a different selection plus a redistribution step: our saliency map picks the keep-set, a REAP-style per-layer floor (p=24) protects the tail, and the 72 evicted experts per layer are folded DERN-style into the survivors rather than discarded. Router, attention and norms are otherwise untouched. No fine-tuning, no distillation.

Built with omnimergekit.

Read this before you load it

Routing is top-8 (num_experts_per_tok: 8) — the base model's native setting, and measured rather than assumed: MBPP-full 0.784 / 0.790 at top-8 against 0.732 / 0.730 at top-10. The Coder sibling bakes top-10; this one does not.
The MTP block is included (mtp.*, in mtp.safetensors, indexed) for speculative decoding.
This checkpoint is text-only. architectures: Qwen3_5MoeForCausalLM, model_type: qwen3_5_moe_text — the base model's vision tower did not survive the redistribution step and is not present here. The Coder sibling's safetensors repo is multimodal (Qwen3_5MoeForConditionalGeneration); this one is not. The vision-<tier> tags on ollama get their vision tower from the Coder mmproj at the GGUF layer, so they are unaffected — but if you need a multimodal safetensors checkpoint, use Coder.

Quantised builds

GGUF (19 tiers, all imatrix, MTP included): Qwen3.6-27B-A3B-CoderX-MTP-GGUF
ollama: ollama run mannix/qwen3.6-27b-a3b-coderx

Evaluation

Q6_K + imatrix, llama.cpp b9700, greedy (temperature 0.0 / top_p 1.0 / top_k 0), one pinned serving geometry per bench, read back from the server log.

Benchmark	CoderX	Coder (184e)	Qwen3.6-35B-A3B (256e)
LiveCodeBench v6 (77q, 24k think / 48k total)	0.727	0.610	0.610
HumanEval+ (164)	0.970	0.951	0.939
MultiPL-E-100 (rs+java+js, 300 completions)	0.887	0.890	0.910

A same-basis repeat of MultiPL-E moved 1.0 pp on batch-scheduling nondeterminism alone, so the 0.33 pp CoderX↔Coder gap is a tie; the 2.33 pp gap to the teacher is real. Per language (CoderX / Coder / 256e): Rust 0.85 / 0.81 / 0.84 · Java 0.89 / 0.90 / 0.93 · JS 0.92 / 0.96 / 0.96.

The full canonical 9-bench suite has not been run on this checkpoint yet; non-code axes (GPQA, MATH-500, IFEval, ARC) are deliberately not quoted here.

Apache-2.0 · research checkpoint.