Détails

Source

TAO Daily

Publication

24 août à 22h02

Lien direct

https://taodaily.io/bittensors-teutonic-sn3-just-jumped-from-10b-to-110b-parameters-in-10-days/

Contenu source (brut)

<div id="bsf_rt_marker"></div> <p class="wp-block-paragraph">Just 10 days after unveiling <a href="https://taodaily.io/teutonic-i-the-10b-model-forged-through-2163-ai-duels/">Teutonic-I</a>, Bittensor founder Jacob Steeves, known as Const, has announced <strong>Teutonic-II-110B</strong>, a permissionless training run for a roughly <strong>110-billion-parameter language model</strong> on the <strong>Teutonic subnet</strong>.</p> <p class="wp-block-paragraph">The project marks a major expansion of Bittensor’s decentralized AI training efforts, scaling the model size by roughly <strong>10×</strong> compared with Teutonic-I.</p> <h2 class="wp-block-heading">A 110B Model Trained by an Open Market</h2> <p class="wp-block-paragraph">Teutonic-II is built around a <strong>sparse Mixture-of-Experts (MoE)</strong> architecture with approximately <strong>110B total parameters and 7.3B activated parameters per token</strong>.</p> <p class="wp-block-paragraph">Unlike conventional AI training runs controlled by a single organization, Teutonic allows anyone with the required compute and technical capabilities to compete.</p> <p class="wp-block-paragraph">Miners are rewarded for producing <strong>measurably better model updates</strong>, rather than simply contributing compute.</p> <ul class="wp-block-list"> <li><strong>~110B total parameters</strong></li> <li><strong>~7.3B activated parameters per token</strong></li> <li><strong>Permissionless participation</strong></li> <li>Model checkpoints and the final dataset planned for open release</li> </ul> <h2 class="wp-block-heading">How Teutonic&#8217;s Training Market Works</h2> <figure class="wp-block-image size-large"><img fetchpriority="high" decoding="async" width="1024" height="480" src="https://taodaily.io/wp-content/uploads/2026/08/image-162-1024x480.png" alt="" class="wp-image-24173" srcset="https://taodaily.io/wp-content/uploads/2026/08/image-162-1024x480.png 1024w, https://taodaily.io/wp-content/uploads/2026/08/image-162-300x141.png 300w, https://taodaily.io/wp-content/uploads/2026/08/image-162-766x359.png 766w, https://taodaily.io/wp-content/uploads/2026/08/image-162.png 1778w" sizes="(max-width: 1024px) 100vw, 1024px" /><figcaption class="wp-element-caption">Teutonic <a href="https://teutonic.ai/">dashboard</a></figcaption></figure> <p class="wp-block-paragraph">Teutonic uses a continuous <strong>king-of-the-hill</strong> system.</p> <p class="wp-block-paragraph">Participants train challenger models independently and submit them for evaluation against the current best checkpoint. Validators then test both models on a held-out dataset using next-token prediction loss.</p> <p class="wp-block-paragraph">A challenger replaces the current model only when it demonstrates a statistically significant improvement.</p> <p class="wp-block-paragraph">This creates a market where participants compete to find better:</p> <ul class="wp-block-list"> <li>Training strategies</li> <li>Data and curricula</li> <li>Model modifications</li> <li>Compute resources</li> <li>Optimization techniques</li> </ul> <p class="wp-block-paragraph">The underlying principle behind Teutonic: <strong>capital and technical ingenuity compete to produce measurable improvements in the model.</strong></p> <h2 class="wp-block-heading">Building on Teutonic-I</h2> <p class="wp-block-paragraph">Teutonic-II follows Teutonic-I, which was launched on August 14 after a 70-day decentralized training competition.</p> <figure class="wp-block-embed is-type-rich is-provider-x wp-block-embed-x"><div class="wp-block-embed__wrapper"> <div class="embed-x"><blockquote class="twitter-tweet" data-width="500" data-dnt="true"><p lang="en" dir="ltr">Today, after 4 months, Subnet 3 (previously <a href="https://x.com/tplr_ai?ref_src=twsrc%5Etfw">@tplr_ai</a>) got retaken by holders through vote.<br><br>We&#39;re celebrating by releasing our first model.<br><br>Teutonic-I: the worlds most performant decentralized LLM.<a href="https://t.co/w0Dt3aef6m">https://t.co/w0Dt3aef6m</a></p>&mdash; const (@const_reborn) <a href="https://x.com/const_reborn/status/2088349997645312205?ref_src=twsrc%5Etfw">August 14, 2026</a></blockquote><script async src="https://platform.x.com/widgets.js" charset="utf-8"></script></div> </div></figure> <p class="wp-block-paragraph">Teutonic-I used a <strong>10B-parameter architecture</strong> and attracted dozens of independent participants competing through successive model improvements.</p> <p class="wp-block-paragraph">Its strongest checkpoint reportedly achieved a <strong>62.28% unweighted average across 11 benchmarks</strong>, outperforming larger models including Quasar Preview 18B and Covenant-72B on the evaluation.</p> <p class="wp-block-paragraph">Teutonic-II now tests whether the same permissionless training mechanism can scale from <strong>10B to 110B parameters</strong>.</p> <h2 class="wp-block-heading">A Reminder on the Sequence of Events</h2> <p class="wp-block-paragraph">Subnet 3 has been through a full ownership change before reaching this point:</p> <ul class="wp-block-list"> <li>The subnet previously ran as Templar under Covenant AI, the team behind the <a href="https://taodaily.io/templars-covenant-72b-becomes-the-largest-decentralized-llm-training-run-ever-completed/">Covenant-72B model</a></li> <li>Token holders reclaimed the subnet through a community vote after four months</li> <li>The new subnet owner then launched Teutonic-I on August 14, 2026, calling it the most performant decentralized language model built to date</li> <li>Teutonic-I ran as a 70-day king-of-the-hill contest between June 2 and August 10, spanning 2,163 duels and 203 accepted kings</li> <li>The strongest checkpoint scored an unweighted average of 62.28% across 11 benchmarks, ahead of Quasar-Preview 18B and Covenant 72B, leading on 8 of the 11 tasks, including a 75.29% score on MMLU</li> <li>About 73 distinct coldkeys took part, with weights published on Hugging Face and the technical report posted at <a href="https://teutonic.ai/paper.html">teutonic.ai/paper.html</a></li> </ul> <p class="wp-block-paragraph">That contest fixed the architecture at 10 billion parameters, leaving miners free to choose their own datasets, optimizers, and hardware.</p> <h2 class="wp-block-heading">One Hundred Ten Billion and Counting</h2> <p class="wp-block-paragraph">Decentralized networks only compete with closed labs by pooling market-priced incentives across scattered compute and talent.</p> <p class="wp-block-paragraph">Teutonic-II is a direct test that an <strong>open, incentive-driven market can train frontier-scale AI without a centralized organization controlling the entire process.</strong></p> <p class="wp-block-paragraph">The aim of Teutonic is to reach as high as <strong>one trillion parameters</strong>, well beyond what any current run has attempted.</p> <p class="wp-block-paragraph">The dashboard at <a href="https://teutonic.ai/">teutonic.ai</a> is already open for anyone with hardware willing to compete for the daily reward pool.</p>