Quasar

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 👇

Just want to confirm our silence is only because we’re deep inside the Quasar cave, building. But as always with Quasar… That’s not it.

Quasar is trending on Hugging Face. For context, HF is the central hub for AI models. Every serious open-source model lives there. We’re already on page two, alongside Xiaomi, Qwen, and Liquid AI and this is just from our small model experiment! The ML community are taking notice. Not just on Bittensor but in the broader open-source world. The open-source community already built MLX and GGUF versions so Quasar can run on MacBooks and local AI setups. We are just beginning.

Bittensor Ecosystem Highlights :: June 8–14, 2026 SUBNET ACHIEVEMENTS [ @chutes_ai - SN64 ] @jon_durbin shared a draft of the Parallax tech report, outlining a MoE training method to reduce per-participant VRAM and FLOPs. > bit.ly/3RYSpFM Chutes also became a launch partner for Respan’s new AI Gateway. > bit.ly/4aAdpJd [ @QuasarModels - SN24 ] Quasar released Quasar-Preview, its first public Quasar model trained on Bittensor: 18B MoE, 2B active and 5M context. > bit.ly/4ekxl3U Quasar is preparing a 10T-token decentralized training run on SN24, starting with a 5T-token phase to produce a stronger checkpoint. > bit.ly/3RZb7gv [ @oroagents - SN15 ] ORO shared its arXiv pre-print, code, data and post-training pipeline for building shopping agents from SN15’s open agentic shopping traces. > bit.ly/43vW3cG [ @webuildscore - SN44 ] Score showed how its 19MB vision model beat larger AI models on object detection while running much faster on CPU. > bit.ly/4os0FtX They also added new comparison pages on @manakoai against ChatGPT, Claude, Roboflow, SAM 3 and other vision AI tools. > bit.ly/4vMuCY7 [ @vidaio_ - SN85 ] Score is partnering with Vidaio to bring vision AI challenges to SN44 and make video archives searchable and actionable. > bit.ly/4ekAbWC [ @yanez__ai - SN54 ] Yanez partnered with Nexartis, an identity and trust infrastructure company, to help verify human, AI model and agent activity across digital transactions. > bit.ly/4eshDUp They also shared in their latest AMA that Yanez has generated $300K+ in 2026 sales, with an active pipeline over $1M and 11 clients. > bit.ly/4v9VUYq [ @trishoolai - SN23 ] Trishool was accepted into Anthropic’s Claude Partner Network. > bit.ly/4uUf5EZ [ @affine_io - SN120 ] AFFINE-XXIX beat the Qwen3-32B baseline on SWE-Rebench, SWE-Multi, HumanEval and MCP-Agent benchmarks, while staying close on BBH. > bit.ly/4xrRA8D [ @VantaTrading - SN8 ] Vanta Trading crossed 2000 users after launching free $1k eval accounts and cutting prices by 55% across all challenges. > bit.ly/3QGKOv5 [ @SwarmSubnet - SN124 ] Swarm announced SOTApilot, an open-source AI drone autonomy model with 95.34% success on its UAV navigation benchmark. > bit.ly/4opsiUi [ @blockmachine_io - SN19 ] Blockmachine launched Ethereum RPC. > bit.ly/4fGe67h [ @TrajectoryRL - SN11 ] TrajectoryRL is expanding SN11’s skill competition from skill packs to miner-submitted finetuned models. > bit.ly/4eoTA91 [ @heydittoai - SN118 ] Ditto reached 1000 users. > bit.ly/4gkB7Na [ @theminos_ai - SN107 ] Minos has run over 37,000 variant-calling evaluations on chromosome 21, with submissions improving by 10.21% on average. > bit.ly/4vO9noW [ @minotaursubnet - SN112 ] Minotaur launched its website and opened beta access to its DEX Aggregator. > bit.ly/447GpEq [ @ReadyAI_ - SN33 ] ReadyAI launched a revenue dashboard showing real-time demand for SN33’s structured data pipeline. > bit.ly/4epth2q [ @say_gm_ - SN28 ] Good Morning published the roadmap for its AI gateway running in a TEE, now live on testnet with mainnet beta next. > bit.ly/4gkzGOQ [ @EndureNet - SN30 ] Endure is integrating @SynthdataCo's forecasts into its DeFi risk engines. > bit.ly/4v5Remt [ @eirel_ai - SN36 ] Eirel released its first product, offering deep research, image generation, web search and agent tools across multiple model families. > bit.ly/4otVg5I [ @adtao_ppcrebel - SN21 ] @dsvfund took an OTC position in the SN21 alpha token. > bit.ly/4eHTs5G SUBNET LAUNCH [ @DeSciClaims - SN111 ] Claims is launching as SN111 to build a claim-evidence graph that turns scientific literature into machine-readable data for AI reasoning. > bit.ly/3SlCUaT PODCASTS & ARTICLES @opentensor Novelty Search hosted by @const_reborn with @zipcodenetwork > bit.ly/3SmDlli @TAO_dot_com Episode 14 with @Carrot_____1 and @KeithSingery > bit.ly/3QCA9S6 @gordonfrayne podcast with @josercaldera from Yanez > bit.ly/4vOufwf @gordonfrayne podcast with @knakamor from Vocence > bit.ly/4493MgY @AltcoinMillie podcast with @MaxScore from Score > bit.ly/3QGDX4L @AltcoinMillie podcast with @zeussubnet > bit.ly/4uBcNKu @TAO_dot_com article “The Impact of Conviction” > bit.ly/4esgBI1

Pushing this harder than ever!

Most large language models are surprisingly forgetful. @QuasarModels (SN24) — built by SILX AI — is trying to fix that at the architecture level: linear-time attention, ~2M-token context, no fragile position embeddings. Memory at the protocol layer. → taoprotocol.org/quasar-sn24-th… $TAO @opentensor

@90daysliquidity We are starting from Quasar-Preview, which we just released a 20B model with 2B active parameters.

@QuasarModels Fantastic news. Is there a stated parameter count or is it being measured in tokens because it’ll be across several different models?

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 👇

@STEEEETY We are focused on architecture and training on a massive number of tokens We are doing a 10T-token run with Quasar 5M context window.

@QuasarModels How is that different to orgo ?

@STEEEETY We simply train Quasar models on a lot of tokens to make them truly smart but the magic is that it will be fully decentralized This has never happened before in the history of decentralized AI

@QuasarModels What does it actually do in dummy language?

TEN TRILLY: Quasar moving toward a 10T-token decentralised training run on Bittensor SN24. Phase 1: 5T tokens to produce a stronger checkpoint. Phase 2: another 5T tokens on top. If successful, this would be one of the largest token-scale training runs in decentralised AI.

A 10T-token run has never happened in the history of decentralized AI But Quasar is special. Quasar has its own unique architecture, and we are nerdy enough to achieve such a task We’ll share how we made this possible and exactly when it begins. Keep an eye out

training.silxinc.com Keep an eye on our account and Discord We will be sharing updates as the training system rolls out It will not become effective immediately. First, we will share the design, roll out the system, run tests, and then fully switch over once everything is ready.

@eliebakouch @teortaxesTex The long-context part was trained on fewer tokens, so we do not claim SOTA long-context performance yet. This is simply our research direction and an experiment.

@teortaxesTex no long context eval btw

What is going on here?

@MooncastOnline This is our experiment and a reveal of our long-context architecture. We have a tech report coming, along with the 10T-token version of the model.

@QuasarModels What is the use case for such a model? It uses a gob of vram, isn't super smart, and has a massive context. That's not a general case model. So what is it for?

Today we’re releasing Quasar-Preview! Our first public proof that the Quasar architecture works at real scale. [ 18B MoE - 2B active / 5M context ] Built with Loop Transformer + Quasar attention Trained on Bittensor through decentralized infrastructure 👇

@iamMrDuncan @Bnonym the goal of this early experiment is to show early learning signals. We are going to train on a 10T-token budget, which will allow us to make stronger comparisons with other highly trained models.

@QuasarModels @Bnonym Can you compare it with Gemma 4 E4B or E2B?

@Bnonym @iamMrDuncan Quasar is only 2B active parameters, This is an early experiment trained on very few tokens.

@iamMrDuncan @QuasarModels Damn, that's some small models. Comparing a 18B with a 4B

@iamMrDuncan This is not about evals or beating other models. Our model is only 2B active parameters, so Quasar is actually comparable to these models. It has also been trained on fewer than 1.5T tokens.

@QuasarModels some strange models you picked to compare benchmarks to.

Let’s fucking go

Going to sleep now. I’ll be uploading a lot of updates and repos for the model and subnet. Have a good night, everyone.