Instructions to use ReliquaryForge/Teutonic-I-Reliquary with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ReliquaryForge/Teutonic-I-Reliquary with Transformers:

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

pipe = pipeline("text-generation", model="ReliquaryForge/Teutonic-I-Reliquary")
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("ReliquaryForge/Teutonic-I-Reliquary")
model = AutoModelForCausalLM.from_pretrained("ReliquaryForge/Teutonic-I-Reliquary", 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 ReliquaryForge/Teutonic-I-Reliquary with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/ReliquaryForge/Teutonic-I-Reliquary

SGLang

How to use ReliquaryForge/Teutonic-I-Reliquary 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 "ReliquaryForge/Teutonic-I-Reliquary" \
    --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": "ReliquaryForge/Teutonic-I-Reliquary",
		"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 "ReliquaryForge/Teutonic-I-Reliquary" \
        --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": "ReliquaryForge/Teutonic-I-Reliquary",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ReliquaryForge/Teutonic-I-Reliquary with Docker Model Runner:
```
docker model run hf.co/ReliquaryForge/Teutonic-I-Reliquary
```

Teutonic-I-Reliquary

A 9B reasoning and coding agent, post-trained entirely on Bittensor, on untrusted GPUs from RTX 3090s to B200s.

On SWE-bench Verified it resolves 43.6% of issues, against 35.1% for Qwen3.5-9B under the same harness and protocol.

Benchmark	Teutonic-I-Reliquary	Qwen3.5-9B
SWE-bench Verified (486 tasks, mini-swe)	43.6	35.1
IFBench	74.7	67.5
AIME 2025	91.2	86.2
LiveCodeBench	62.0	57.0
AIME 2026	89.6	86.7
MMLU-Pro	79.1	80.3
TMax (terminal tasks)	74.8	82.2
GPQA Diamond	68.1	81.6

pass@1 (%). Both models were evaluated on the Reliquary network itself, with the same harness, budgets and sampling. Protocol below.

How it was trained

Stage 1: SFT on data generated by miners

Before RL, Teutonic-I was fine-tuned on about 4.5B tokens generated by Reliquary miners, including 24,500 multi-turn agentic episodes. On our internal evaluations, SFT took it from 0% to 34% on real-world bug fixing and from 0% to 26% on tool use, close to Qwen3.5-9B on both.

Stage 2: RL on an open network

teutonic-rl-async-v1, October 22 to November 2, 2026.

158 optimizer steps over 11 days, 15.7B trained tokens. No step skipped, no gradient clipped.
Open to any GPU. Anyone with a GPU could join; miners came with 11+ GPU models, consumer and datacenter cards side by side.
Fully asynchronous. The trainer published a new checkpoint after every step, and no miner ever waited for it.
7 environments. SWE tasks, terminal tasks, competitive programming, hard math, DAPO math, tool-use worlds and instruction following, with agentic episodes up to 48 turns and 64k tokens.
Untrusted by design. Miners send tokens only. The network computes every reward and verifies every rollout.

On TMax, a held-out split of terminal tasks, the model went from 17.2 to 74.6 over the run (evaluated on the network at each checkpoint). Every step is on the run dashboard.

Heterogeneous, untrusted, asynchronous. Each one alone is hard for RL; together, the usual fixes stop working. This run had all three. How: Gumbel substitution, technical report coming soon.

Post-trained entirely on Bittensor

Stage	Where
Base model	Teutonic-I, Teutonic (Subnet 3)
SFT data: 4.5B tokens, including 24,500 agentic episodes	Generated by Reliquary miners (Subnet 81)
SFT and RL trainer compute	Lium (Subnet 51)
Reinforcement-learning rollouts	Reliquary miners
Sandboxes for SWE and terminal tasks	Targon (Subnet 4)
Evaluation: every benchmark sample	Generated and verified by Reliquary miners

Lineage details for the base and the SFT stage are in the SFT model card.

Usage

Same architecture and chat template as Qwen3.5-9B (Qwen3_5ForCausalLM), thinking mode.

vllm serve ReliquaryForge/Teutonic-I-Reliquary --max-model-len 65536

Recommended sampling, as in our evaluations: temperature 1.0, top-p 0.95, top-k 20.

from transformers import AutoModelForCausalLM, AutoTokenizer

name = "ReliquaryForge/Teutonic-I-Reliquary"
tok = AutoTokenizer.from_pretrained(name)
model = AutoModelForCausalLM.from_pretrained(name, torch_dtype="auto", device_map="auto")
msgs = [{"role": "user", "content": "Prove that the square root of 2 is irrational."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=8192, do_sample=True, temperature=1.0, top_p=0.95, top_k=20)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

Evaluation protocol

Generated on the network. Every sample was produced by Reliquary miners and verified by the network. A sample that fails verification is refused, never corrected. A missing or truncated sample counts as a failure.
Sampling: temperature 1.0, top-p 0.95, top-k 20, thinking mode, for both models.
Budgets: 64k tokens for AIME, GPQA Diamond and LiveCodeBench; 32k for IFBench and MMLU-Pro. SWE-bench Verified (mini-swe harness) and TMax: up to 48 turns, 16k tokens per turn.
Sizes: AIME 2025 and 2026, 30 problems × 8 samples; GPQA Diamond 198 × 4; LiveCodeBench 300 × 2; IFBench 300 × 2; MMLU-Pro 500 × 2; SWE-bench Verified 486 × 2 (the 500 tasks minus 7 that need network access and 7 whose reference patch fails in our sandbox); TMax 200 × 4.
Held out. TMax is scored on a split never used in training.

License

Apache-2.0, as its base models.

Downloads last month: -

Safetensors

Model size

9B params

Tensor type

BF16

Model tree for ReliquaryForge/Teutonic-I-Reliquary

Base model

Qwen/Qwen3.5-9B-Base

Finetuned

ReliquaryForge/Teutonic-I-Reliquary-SFT

Finetuned

(2)

this model