SN 102 @ConnitoAI
ESFT is a fine-tuning optimization technique designed for Large Language Models (LLMs) that use a Mixture-of-Experts (MoE) architecture !!! 4X Fasters at the same accuray
MoE !!!!
$TAO Break on Through ⬇️⬇️
ConnitoAI (Bittensor Subnet 102): Major Training Breakthrough Just Dropped
This is big.⬇️⬇️
https://t.co/h24Rr6CRSo
@ConnitoAI just announced a serious technical win: they’re now training with 5× less memory and 4× faster than previous methods using their new In-place ESFT (Efficient Sparse Fine-Tuning).
The Numbers (June 2026 Status):
• Active parameters during training: 3.11B (only ~20% of the full 15.7B model)
• Peak GPU memory: 21 GB (down 83%)
• Training time: 3h 07m (4× faster)
• Accuracy: Maintained at 39.5% on GSM8K-CoT math benchmark
What It Means
Instead of loading the entire Mixture-of-Experts model on GPU, miners now train only a small active shard of specialists in-place, park the rest on CPU, and merge updates seamlessly.
This makes decentralized training dramatically more efficient and accessible on affordable hardware (single A6000 GPUs).
Bottom line: ConnitoAI just lowered the cost and raised the speed of specialized model training across the network. This directly strengthens the decentralized training factory thesis and accelerates the flywheel for custom expert modules.
Real execution. Real iteration. The MoE beast is getting stronger.
RELEASE THE KRAKEN. 🐙⚡️
$TAO $SN102 $SOL
#ConnitoAI #SN102 #Bittensor #DecentralizedAI
DYOR. The training wars just leveled up. 🚀
SN102 continues to show strong technical progress.
Recent dashboard data shows validation loss improving from ~0.20 to ~0.14, while competition among miners remains intense with ~225-255 active miners. Meanwhile, Connito’s latest research claims its in-place ESFT approach achieves performance comparable to classic fine-tuning while using ~80% less GPU memory and training ~4x faster.
The technology is working. The next milestone is proving real customer demand.
@connitoai #ConnitoAI #SN102 #Bittensor #TAO #DecentralizedAI #AI
The technical results look promising. The next milestone is proving demand: real customers, real workloads, and real revenue. If that happens, SN102 could be something special.
@connitoai#ConnitoAI#SN102#Bittensor#TAO
- My personal opinion.
@const_reborn
you deserve thanks, and so does everyone who helped build this ecosystem.
What started as a network has become a catalyst for change. Not only financially for many people, but in ways that will continue to touch lives across countless fields.
AI is no longer a future concept; it’s a reality. And what the different Bittensor subnets are building today may become the foundation for ideas, tools, and services that make life easier for millions tomorrow.
I hope to see this vision continue to grow and reach its full potential.
#TAO #Bittensor
The #BV7X token has crossed $10M in trading volume since launch. 🔥
4 months on @base — Bitcoin intelligence put on-chain.
Every call attested, every trade verifiable. The oracle is just warming up. 🧠⚡
https://t.co/tM1v8X59pp
This week at Vocence:
• Continued improving StyleTrajTTS expressive speech generation
• Reduced latency across our inference pipeline
• Expanded multilingual voice evaluations
• Testing longer-form conversational consistency
• Preparing new enterprise voice agent
The future of voice AI isn't just sounding human.
It's maintaining behavior, personality, and expression throughout an entire conversation
Public Disclosure: First SOMA OTC Transaction
We're excited to share that 25,025 SOMA subnet owner tokens have been acquired by @const_reborn through an OTC transaction.
This is our first OTC deal, and it's a strong signal of confidence in what we're building. The proceeds will directly accelerate development and bring our product closer to public release.
Thank you, Const. This vote of confidence means a lot to the entire team and motivates us to keep building the future of SOMA.
We’re just 1 hour away from the Vocence (Subnet 78) AMA with @knakamor
She’ll be diving deep into what the team is building at @vocence_bt , the decentralized voice intelligence layer on Bittensor.
Location: Subnet Summer telegram Community
Join here: https://t.co/pa9NG7Q4vp
Bring your best questions about:
• Miner & validator mechanics
• Text-to-Speech, Speech-to-Text, Voice Cloning & more
• Token incentives and roadmap
• How they’re pushing open-source voice AI forward
Don’t miss it, this is your chance to ask the team directly.
Update:
Daemon needed deeper forensics tooling, so we’re bringing RicoMaps into the ecosystem.
Built by our own @nullxnothing, @RicoxMaps is a Solana on-chain intelligence platform for mapping wallet and token relationships in real time.
Paste a wallet or token address and uncover funding chains, shared funders, holder clusters, first-block snipers, Jito bundle activity, wallet profiles, and risk signals across the network.
The goal is simple:
Make Solana investigations visual, fast, and local.
RicoMaps gives builders, traders, and communities a clearer way to follow the money, spot coordinated behavior, and understand what’s happening behind the chart.
$REPPO is the most asymmetric bet on BASE, and no other project comes close to it in terms of fundamentals.
They are the data infra for AI, and secured more than $20M in investment.
They will announce a very big partnership too, so I think these levels will be a dream soon.
Ask Daemon is live.
We wanted a Daemon tool holders could actually use, even if they’re not devs.
Paste a wallet, token, tx, repo, or idea and get a clean Solana read back. Simple for supporters, useful for builders.
Daemon holders get access to deeper onchain reads.
@meterflowsol & @keycardsol holders are also granted access.
Try this out now:
https://t.co/TQF0wA7Ej8
wake up @toly:
@DaemonTerminal is spawning a new ecosystem for projects launched on @solana using their IDE for devs
imo $daemon @ $160k is stupidly undervalued… expecting them to win something at the Solana frontier hackathon
4vpf4qNtNVkvz2dm5qL2mT6jBXH9gDY8qH2QsHN5pump
DAEMON is expanding.
We built first for developers, but our community is bigger than just developers.
Founders, holders, creators, traders, teams, and supporters should all have a reason to open DAEMON and actually use it.
Today, we’re pushing an update that makes DAEMON more accessible to everyone.
Not less powerful.
Just easier to use.
Introducing Daemon Access, powered by @keycardsol .
With Keycard, we can now gate Daemon subscriptions by real token holdings.
That means access is not limited to one community.
$DAEMON, $KEYCARD, and $MFLOW (@meterflowsol) holders will all be eligible to receive DAEMON Access.
One ecosystem. Shared utility. Real holder benefits.