Quasar

Bittensor subnet built to crush the long-context barrier | SN24 |
Owners : @yousseffarahat0 and @troyquasar
Backed by @const_reborn

[Quasar @QuasarModels]

Quasar is entering its next chapter on Bittensor SN24.

We are moving toward a 10T-token decentralized training run.

The idea is simple Quasar Models needs more useful training, not just bigger parameter counts.

Real model quality comes from tokens, data quality, training direction, and the ability to keep improving checkpoint by checkpoint.

This run starts with a 5T-token phase to produce a stronger checkpoint, then continues into another 5T-token phase, reaching 10T total trained tokens.

SN24 will set the direction:

the starting checkpoint, the data, the training recipe, and the evaluation system.

Miners become the extra compute layer.

They help Quasar train faster, improve continuously, and move forward together.

this would be the largest token-scale training runs ever attempted in decentralized AI.

This is how we scale Quasar.

Help us train it 👇

[Quasar @QuasarModels]

https://t.co/pS8EAhRll4

[Quasar @QuasarModels]

Official Quasar Response Regarding Quasar-Preview

We have published a detailed document addressing the concerns, technical claims, and questions raised regarding the Quasar-Preview model.

Eyad Gomaa, Quasar’s CTO and co-founder, led the model’s pretraining and research efforts. In this document, he explains how Quasar-Preview was developed, its relationship to Ling, the checkpoint mismatch, the initialization and training process, and the evidence supporting the intended model.

We acknowledge that our original documentation did not explain Quasar-Preview’s lineage, retained components, initialization, and training process clearly enough. We also failed to ensure that the checkpoint publicly available on Hugging Face accurately represented the latest state of the model. These failures created legitimate confusion and gave the community valid reasons to ask serious questions.

As the owner of SN24 and the person responsible for Quasar’s technical direction, Eyad takes responsibility for these documentation, communication, and checkpoint-management failures.

However, these failures do not support the broader conclusions that Quasar was simply an unchanged copy of Ling, that its training was fabricated, that attribution was intentionally concealed, or that the team intended to mislead the community, take project funds, dump, and disappear.

Those conclusions were drawn before the intended checkpoint, complete training history, attribution record, and supporting technical evidence had been fully reviewed.

A clearer release from us and a more complete technical review before definitive public conclusions were reached could have prevented much of the resulting confusion and harm.

We sincerely apologize to the Bittensor community, Quasar holders, and everyone whose trust was affected by the mistakes on our side. We are publishing the full response because we also believe it is important to correct the allegations that are not supported by the evidence and allow Quasar to be evaluated on the complete technical record

https://t.co/uqL3UtJYuv

[Quasar @QuasarModels]

Quasar is making a major long-term commitment to SN24 by fully locking approximately 130,000 Alpha in the owner wallet.

We were among the earliest teams to test, support, and believe in this conviction.

We are not here for short-term extraction. We are here to build, contribute, and remain fully aligned with the long-term success of SN24.

[Quasar @QuasarModels]

Quasar’s biggest update yet and a major milestone for decentralized AI training is coming to Novelty Search

📅 Friday, July 31
⏰ 12:00 AM EEST (GMT+3)

We’ll reveal upcoming Quasar updates, exciting numbers, technical breakthroughs, and the massive progress we’ve made.

[siamkidd @SiamKidd]

The @QuasarModels team has paid $0 in launching SN27 @OrionSILX They came to us (we owned 27) with a deal and we accepted their terms (zero haggling) in order to alleviate build friction.

There's just a diminishing owner key schedule for us. That's it. The team are very smart and Orion and Quasar compliment each other.

[Quasar @QuasarModels]

Quasar training update 🚀

We are only around 10% into the current training run, and we are already seeing clear benchmark movement.

Old Quasar → Since Training

MMLU: 68.40% → 69.01% (+0.61)
MMLU-Pro: 33.20% → 33.72% (+0.52)
GPQA: 25.60% → 26.01% (+0.41)

ARC Challenge: 63.00% → 64.10% (+1.10)
ARC Easy: 80.10% → 83.96% (+3.86)

PIQA: 81.90% → 81.50% (-0.40)
HellaSwag: 74.00% → 74.79% (+0.79)
OpenBookQA: 47.00% → 46.40% (-0.60)
MATH-500: 71.40% → 72.26% (+0.86)

It has only been two weeks since we started this training run, and we are still running with relatively low compute.

As we add more compute support, we expect the training to scale further and the results to keep improving.

Quasar is just getting started.

[bitstarter @bitstarterAI]

We launched @QuasarModels January. Six months later, they've acquired SN27 and turned it into @OrionSILX, a data engine built to fuel their models and the wider Bittensor ecosystem.

The community gets it: vertical integration, flywheel spinning. Huge. 🛰️

Better data → better models → better data. This is what building as a network looks like.🔥

[Quasar @QuasarModels]

Your Compute. Your AI. Train Quasar. Own the Future 💫

[siamkidd @SiamKidd]

New subnet birth... SN27 @OrionSILX https://t.co/7Bp1tsz8za

[Punisher ττ ✠ @CryptoZPunisher]

Bittensor >> New Subnet >> $dTAO

Subnet 27: Orion
@OrionSILX

[Quasar @QuasarModels]

https://t.co/jOoLAulwNY

[Orion @OrionSILX]

🌌 Introducing Orion, SN27

Orion is a decentralized data subnet built by the same team behind @QuasarModels

Our incentive mechanism is designed specifically to discover, generate, and curate the highest-quality data possible for training AI models.

By using miners, we can reduce the cost of producing high-quality datasets from approximately $100,000 to as little as $10,000, while enabling miners across the network to help discover billions of valuable training tokens in a decentralized and cost-efficient way.

We are starting by providing high-quality data for SN24 Quasar and SN3, helping both networks scale their AI training further.

But Orion will not stop at Bittensor.

We are building Orion AI, a product that will provide high-quality training data to AI developers and companies around the world. We plan to begin rolling out the beta next month.

We are also excited to welcome @MarkCreaser and @SiamKidd as investors, and we are grateful to @const_reborn for advising us and helping shape Orion’s incentive mechanism.

We will begin releasing the incentive mechanism, sharing the code, and publishing regular development updates through our Discord.

Keep an eye on Orion.This is only the beginning. 🚀

[Marcel | TAO Research @marceltaores]

This isn't just another model update.

It's one of the clearest visions I've read for the future of decentralized AI.

Most labs are racing to build smarter models.

@QuasarModels is trying to solve what may be AI's biggest unsolved bottleneck:

Memory.

Not RAM. Not context size on paper.

Real long-context intelligence.

The ability to reason across massive working contexts without forgetting, compressing, or losing coherence over time.

That changes everything.

Today's AI agents often rely on RAG, vector databases, summaries, and external memory to compensate for what the model cannot retain. Those techniques work—but they're still workarounds.

@QuasarModels 's thesis is different:

Solve the problem at the model level, not at the product layer.

That's why they're building the model before the product.

As they put it:

> "The model is the product."

Instead of chasing parameter counts, they're focusing on architecture, training quality, and data quality.

Their current MoE model activates only a fraction of its parameters while maintaining strong long-context performance—around 90% at 100K tokens and 84% at 1M tokens according to their internal benchmarks.

But the ambition goes much further.

• A fully decentralized 10 trillion-token training run.
• Scaling from 20B → 40B → 100B parameters.
• Building a true frontier model on decentralized infrastructure.

To me, the biggest takeaway isn't another benchmark.

It's the vision.

@QuasarModels isn't trying to prove that decentralized AI is possible.

They're trying to prove that decentralized AI can lead.

If they succeed, it won't just be a win for Subnet 24.

It could redefine what's possible for the entire bittensor:native ecosystem.

[Quasar @QuasarModels]

https://t.co/MnToIOfFP5

[Quasar @QuasarModels]

@xtaohq Its already up :

https://t.co/FI1eGx3DFH

[TAO Planet @TAO_Planet_]

Why I believe @QuasarModels (SN24) is one of the most overlooked subnets in Bittensor? 🧵

Everyone is chasing the next hot subnet.

Meanwhile, some of the biggest opportunities are quietly being built in the background.

For me, Quasar (SN24) is one of them.

Why?

• It focuses on one of AI's biggest bottlenecks: high-quality data.

• Every AI model is only as good as the data it learns from. Better data leads to better models.

• Quasar is building a decentralized marketplace that incentivizes the creation and curation of valuable datasets.

• While attention is flowing elsewhere, the team continues to build.

If Bittensor succeeds, infrastructure subnets that solve fundamental problems could become some of the most valuable parts of the ecosystem.

That's why I see Quasar as a sleeping giant.

Not because of hype.

Because of the problem it's trying to solve.

$TAO

#Bittensor #TAO #SN24 #Quasar #DeAI

[Gordon Frayne @gordonfrayne]

Quasar (SN24): Long-Context Intelligence on Bittensor TAO https://t.co/iz22Lv7Q1V

[Gordon Frayne @gordonfrayne]

I'll be going live on X and Youtube with Gomma from @QuasarModels tomorrow morning to dive into all the latest updates on their long-context intelligence being built on Bittensor.

See you at 11am EDT tomorrow!

[Openτensor Foundaτion @opentensor]

Bittensor Ecosystem Highlights :: June 22–28, 2026

[ @chutes_ai - SN64 ]
Chutes is now a built-in provider in @TypingMindApp. They also added a VS Code extension for running Chutes models as coding agents.
> https://t.co/nfsjrasMbt
> https://t.co/McLwZF6Ib5

@jon_durbin also shared new Parallax details and a diffusion-draft pipeline showing 30–50% real-world decode performance gains on Gemma-4 31B and Qwen3-6-27B.
> https://t.co/G76pvrVs2U
> https://t.co/DEKJts11LJ

[ @Ninja_Subnet - SN66 ]
Ninja launched Katana, an agentic IDE and Bittensor-native workspace for research, mining workflows, code shipping and validator debugging.
> https://t.co/Cl8JWujHKS

[ @webuildscore - SN44 ]
Score announced Satori 1.0, its first ~2B VLM, distilled from larger models to handle nine vision primitives, and shared plans to train future versions on SN44 using Teutonic SN3's mechanism.
> https://t.co/Myxtp0lQv7
> https://t.co/nso0VtVin0

[ @bitsecai - SN60 ]
Bitsec outperformed Fable 5 on a security audit, finding 160+ vulnerabilities including five criticals and ten highs that Fable 5 missed.
> https://t.co/oDaBDAcc33

[ @lium_io - SN51 ]
Lium completed a 2,500 TAO buyback and burn into subnet 51, funded entirely by revenue from GPU credit purchases.
> https://t.co/XkTo3Q4LD2

[ @QuasarModels - SN24 ]
Quasar launched its 10T-token incentive mechanism, starting with a 5T-token target and support from Gradients on post-training and RL.
> https://t.co/8jnEDOB9UH
> https://t.co/OURw11uSiK

[ @theminos_ai - SN107 ]
Minos launched MinosVM 2.0 on Targon with native AI assistant support for miners.
> https://t.co/shcarbg1tT

[ @affine_io - SN120 ]
Affine moved its base model to Qwen3.6-35B-A3B, with new champions pushing benchmark scores higher across Memory, NavWorld, SWE and Terminal.
> https://t.co/EWsIuZL2pG

[ @heydittoai - SN118 ]
Ditto launched its mobile app on Google Play and the Apple App Store.
> https://t.co/jMe92QwNnZ

[ @b1m_ai - SN105 ]
Beam successfully transferred 50GB in ~51 seconds across multiple sources and R2 destinations.
> https://t.co/PPnLGzVfAk

[ @Apex_SN1 - SN1 ]
Apex redesigned its website with dedicated agent profiles to showcase user contributions, peer activity and status across SN1.
> https://t.co/qBCGs1DMPy

[Marcel | TAO Research @marceltaores]

A Bittensor subnet with just a $10M market cap is already making waves far beyond the Bittensor ecosystem.

@QuasarModels 's model is now trending on Hugging Face, appearing alongside industry giants like Xiaomi and Qwen.

For those unfamiliar, Hugging Face is the world's leading open platform for AI models. It's where researchers, developers, and companies discover, share, and deploy state-of-the-art AI. Trending there means your work is gaining real visibility and recognition from the global AI community.

Most people have never heard of Hugging Face, which is why it's easy for investors to overlook what this means.

But for AI engineers, researchers, developers, and miners, Hugging Face is one of the most important platforms in the industry. It's where cutting-edge models are discovered, tested, and shared with the global AI community.

By the time this level of attention reaches mainstream social media, crypto influencers, and retail investors, the rocket will likely have already left the launch pad.

#Sn24 #TAO