Instructions to use h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3 with Transformers:

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

pipe = pipeline("text-generation", model="h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3")
model = AutoModelForCausalLM.from_pretrained("h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3", 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 h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3

SGLang

How to use h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3 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 "h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3" \
    --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": "h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3",
		"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 "h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3" \
        --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": "h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3 with Docker Model Runner:
```
docker model run hf.co/h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3
```

Sovereign Judge 35B — BF16 Merged

This model is not a general-purpose chat assistant. It is a specialised quality-assurance gate designed to evaluate, score, and arbitrate the outputs of a Multi-Expert Orchestration (MoE) pipeline before they reach the user.

Full BF16 merged model of the Sovereign Judge, created by merging Qwen/Qwen3-30B-A3B with the LoRA v3 adapter. This is the reference artifact for arbitrary requantization (GGUF, GPTQ, AWQ, …). For deployment, prefer one of the quantised variants below.

For the full technical documentation — output schema, paraconsistent logic, conflict arbitration protocol, and pipeline role — see the GGUF repository.

Model variants

Variant	Repository	Size	Recommended for
BF16 Merged (this repo)	`h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3`	~65 GB	Requantization, research
GGUF Q4_K_M	h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3-GGUF	~21 GB	Ollama / llama.cpp
W4A16 GPTQ	h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3-W4A16-GPTQ	~19 GB	vLLM
LoRA Adapter	h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3-LoRA	~83 MB	Further fine-tuning

What this model does

The Sovereign Judge is the last quality barrier before a synthesised answer leaves the MoE Sovereign pipeline. Its three core responsibilities:

Responsibility	Description
Factual correctness	Detects unsupported claims against retrieved source material
Internal consistency	Identifies logical inconsistencies across parallel expert outputs
Conflict arbitration	Resolves contradictions between domain experts using Belnap-Dunn four-valued logic without logically collapsing on them

When called for conflict arbitration, the model produces structured output:

<conflict_map>
{
  "points_of_dispute": [
    {
      "point": "...",
      "evidence_a": "...",
      "evidence_b": "...",
      "bilattice_value": "<T | F | I | U>"
    }
  ]
}
</conflict_map>
VERDICT: <A | B | SYNTHESIS> — <rationale>

The I (Inconsistent) bilattice value records a genuine contradiction without forcing a premature resolution — the core of paraconsistent reasoning (de Vries 2007, arXiv:0707.2161).

Model architecture

Property	Value
Architecture	`Qwen3_5MoeForCausalLM`
Total parameters	~35B
Active parameters per token	~3B (A3B — 8 of 256 experts active)
Experts per layer	256
Active experts per token	8
Hidden layers	40
Hidden size	2 048
Attention heads	16 (KV-heads: 2, GQA)
Context window	262 144 tokens
Vocabulary	248 320
Weight precision	BF16

Training details

Parameter	Value
Base model	`Qwen/Qwen3-30B-A3B`
Fine-tuning method	LoRA SFT v3 (r=64, α=128)
Dataset	`paraconsistent_large.jsonl` — 90 103 examples, Alpaca format
Dataset languages	German and English
Epochs	3
Total training steps	2 112
Effective batch size	128 (2 per GPU × 8 grad accum × 8 GPUs)
Learning rate	2 × 10⁻⁴
Framework	TRL + DeepSpeed ZeRO-3
Hardware	8 × AMD MI250X (LUMI-G, CSC Finland)

Final training metrics (epoch 3)

Metric	Value
Training loss	0.3032
Mean token accuracy	86.72 %
Entropy	0.4447

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

arbitration_prompt = (
    "Two experts in 'medical_consult' produced conflicting claims.\n"
    "Perform a paraconsistent bilattice evaluation (Belnap-Dunn logic: T, F, I, U) "
    "to arbitrate the dispute.\n\n"
    "CLAIM A:\n<expert A output>\n\n"
    "CLAIM B:\n<expert B output>\n\n"
    "Format your response with a JSON conflict map inside XML tags and a final verdict:\n"
    "<conflict_map>\n{ \"points_of_dispute\": [...] }\n</conflict_map>\n"
    "VERDICT: <A|B|SYNTHESIS> — <rationale>"
)

inputs = tokenizer(arbitration_prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.1)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

VRAM requirements (BF16)

Configuration	Minimum VRAM
1× GPU	80 GB (A100 80G)
2× GPU	2× 48 GB
4× GPU	4× 24 GB

For consumer hardware use the GGUF Q4_K_M variant (~21 GB, Ollama).

Requantization from this BF16

# GGUF — requires --no-mtp flag (mtp_num_hidden_layers is set in config.json
# but no MTP tensors exist in the merged model; without the flag block_count=41
# and blk.40.attn_norm.weight goes missing)
python convert_hf_to_gguf.py h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3 \
    --outtype q4_k_m --no-mtp

# AWQ
python -m awq.entry --model_path h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3 \
    --w_bit 4 --q_group_size 128

# GPTQ — see W4A16-GPTQ repo for the llmcompressor recipe used

License

Based on Qwen/Qwen3-30B-A3B — subject to the Qwen License Agreement.

Citation

@misc{sovereign-judge-2026,
  title  = {Sovereign Judge 35B — BF16 Merged},
  author = {Philipp Horn},
  year   = {2026},
  note   = {LoRA v3 fine-tune of Qwen3-MoE-35B for paraconsistent quality arbitration,
            part of the MoE Sovereign project},
  url    = {https://huggingface.co/h3rb3rn/Qwen3-MoE-35B-Sovereign-Judge-v3}
}