SN3 is now training Quasar models.
This time, weβre starting with a 10B model built on the new Quasar architecture an architecture we believe represents the future of LLMs, especially for long-context reasoning and memory.
Quasar is designed to push models beyond short-window intelligence and into systems that can truly handle massive context, retain structure, and reason across long sequences.
Weβre calling on miners to help us take the best out of this architecture, improve it, optimize it, train it harder, and push Quasar models forward round after round.
Letβs build state of the art.
Unlock Bittensor's hidden Conviction signals via the FREE Tensia Bot from @TensiaFDN . The bot makes on-chain capital commitment visible. Real-time alerts & insights.
@Taoalertstrackerbot (telegram)
https://t.co/EurGPSfu7k (Website)
@gregoriusthe3rd@TroyQuasar@TheTNetHunter@adaption_ai Thanks for the insight fren, a great analogy. My original comment was about complimentation. i couldnt see the wood for the trees but now understand alot more on how @QuasarModels will achieve this.They dont have to build on top (my confusion).
@TroyQuasar@TheTNetHunter@adaption_ai Just think I've found the part I was missing.
So a large model say Orion 100B could call Quasar via tool calling for long context?
@TroyQuasar@TheTNetHunter@adaption_ai But my point is what you are building could compliment models built on Bittensor by giving said models (hopefully soon) SOTA context. If I'm wrong,lesson learned I will do more research.
Orion-100B was made possible by a series of advances:
- The creation and utilization of ResBM, currently the state-of-the-art (SOTA) technique for lossless activation compression in LLM training,
-A custom, fault-tolerant peer-to-peer network protocol that optimizes throughput & latency across heterogeneous GPU nodes.
-Reliable distributed variable synchronization.
These advancements have enabled IOTA to increase its MFU capabilities by an order of magnitude when compared to previous results, and support training across many pipeline stages without significant costs to net throughput.
Orion-100B from @IOTA_SN9 just shattered the ceiling on decentralized AI training
A full 100-billion-parameter model trained across 48 single A100 GPUs in 5 different US datacenters, coordinated entirely over the open internet via Bittensor's Subnet 9. The team achieved 30% model FLOP utilization β roughly 65% of datacenter training speed β while using distributed hardware that costs a fraction as much, thanks to their breakthrough ResBM compression that shrinks activation data by 64Γ. A single contributor with one GPU can now participate in frontier-scale training, and this system scaled 67Γ in model size in just one month. The era of truly permissionless AI is no longer theoretical β it's training a 100B model right now.
Today, we are launching the first stage of Project Orion.
Our early pre-training run of Orion-100B achieves upward of 65% of data-center training efficiency on hardware costing a fraction of the price.
Orion-100B is the first proof point for a simple idea: that underutilized compute around the world can be turned into frontier training capacity.
We believe that this work presents, for the first time, an economically compelling case for training large models using distributed approaches.
MVTRX Teaser Video #2
Hedge funds had this for decades. You didn't.
GenTRX changes that.
AI trained on the full history of the exchange, built collectively by the market, accessible to every trader.