Reliquary (SN81) Post-Trained a 9B Model That Beats Qwen3.5-9B on Five Benchmarks

Teutonic-I-Reliquary, a nine-billion-parameter AI model built for coding and reasoning, went through its final training on Bittensor through Reliquary (SN81).

It scored 43.6% on SWE-bench Verified, a benchmark that tests whether AI can fix real software issues. Qwen3.5-9B scored 35.1% under the same test conditions. Teutonic-I-Reliquary led on five of the eight benchmarks.

Reliquary (SN81) is a virtual arena that lets miners use their own GPUs to improve an existing AI model, a process called post-training.

Miners contribute examples of how to solve tasks, run the model to generate fresh attempts, and evaluate its work. Those examples and results help shape what the model learns during training.

For this run, Teutonic (SN3) provided the starting model, Lium (SN51) supplied the computing power for training, and Targon (SN4) ran the isolated environments where AI agents could execute code.

On Reliquary, 161 miners contributed over 11 days using more than 11 GPU models, from RTX 3090s to B200s. The miners supplied 4.5 billion tokens of fine-tuning examples. Across the full run, training processed 15.7 billion tokens over 158 steps.

Different GPUs can produce slightly different results when running the same model, making it harder to train reliably across independently operated machines.

Reliquary used a technique called Gumbel substitution to keep reinforcement learning stable despite those differences. The network also verified every attempt generated for that stage of training.

During reinforcement learning, the model learned from how its attempts scored. Its results improved across all eight benchmarks. SWE-bench Verified rose from 17.3% to 43.6%, LiveCodeBench climbed from 40.7% to 62.0%, and AIME 2025, a maths reasoning benchmark, went from 73.3% to 91.2%.

The model weights are available under the Apache 2.0 licence, so developers can download and adapt the model. Reliquary also released the full training record, giving researchers a way to inspect how the model was improved across independently operated GPUs.

➛ Access Teutonic-I-Reliquary on Hugging Face and Its Training Dashboard.