TensorQ

Autonomous Alpha Trading on Bittensor

[TensorQ @TensorqTao]

TensorQ is now live on BASE.

No black box, no guesswork. Just on-chain data and a system that learns.

$TENSORQ: 0xC3EEC1f0CbA5775F0F9f0B7f9F3dB30770F844E8

Explore the TensorQLLM: https://t.co/MF6n5t66iY
Documentation: https://t.co/xVSRXkHifj

[TensorQ @TensorqTao]

QLLM-8B fine-tuning rounds now pay 2x the standard reward rate.

Workers download the Llama 3.1 8B base + latest LoRA adapter, fine-tune on conversation shards, and submit their adapter weights. Best adapter per round gets promoted.

16GB+ of VRAM required. Two commands to start:

pip install torch transformers peft trl datasets bitsandbytes python https://t.co/H9xPfK87Uv --wallet 0xYourAddress

https://t.co/qC8PPDgHBj

[TensorQ @TensorqTao]

Most fine-tuned trading models are trained on price history or synthetic data.

QLLM-8B is trained on 720 real Claude Opus conversations — actual subnet analysis sessions where the model reasoned through emissions, stake flows, market regimes, and produced structured trade decisions.

The training data is what the agent actually thinks. Not a simulation of it.

https://t.co/qC8PPDgHBj

[TensorQ @TensorqTao]

QLLM-8B under the hood:

Base: Llama 3.1 8B Instruct (frozen)
Method: QLoRA 4-bit NF4, LoRA rank 64, alpha 128
Trainable params: 167M out of 8B total
Train loss: 0.069 · Token accuracy: 98%
Adapter size: 657MB
Training time: 38 min on H100

Runs on a single RTX 4090 at $0.45/hr. Consumer GPU friendly: 3090 or 4070+ with 16GB VRAM.

https://t.co/qC8PPDgHBj

[TensorQ @TensorqTao]

QLLM just got a major upgrade.

The SubnetForecaster is out. QLLM-8B is in — a QLoRA fine-tune of Llama 3.1 8B Instruct, trained on 720 real Claude Opus subnet trading conversations.

Same data moat. Completely different architecture. Outputs structured JSON buy/sell/hold recommendations with full reasoning — the same format Claude Opus uses.

https://t.co/qC8PPDgHBj

[TensorQ @TensorqTao]

QLLM just got a major upgrade.

The SubnetForecaster is out. QLLM-8B is in — a QLoRA fine-tune of Llama 3.1 8B Instruct, trained on 720 real Claude Opus subnet trading conversations.

Same data moat. Completely different architecture. Outputs structured JSON buy/sell/hold recommendations with full reasoning — the same format Claude Opus uses.

https://t.co/qC8PPDgHBj

[TensorQ @TensorqTao]

Recent fixes made to the TensorQ Infra:

• Monte Carlo backtester was mathematically broken (shuffled sums are invariant). Fixed to use equity curve Sharpe.
• Strategies now must beat TAO buy & hold to pass — no more "good Sharpe, worse than holding."
• QLLM predictions tracked and resolved at 1h/6h/24h — accuracy chart live on the model page.
• Swarm evolution fixed — 32 generations of zero trades because confidence thresholds were unreachable.
• Community training upgraded from numerical model to full QLLM-8B QLoRA fine-tuning.

Everything is live. Everything auto-refreshes. The agents don't sleep.

https://t.co/w4Bb18Y7Ot

[TensorQ @TensorqTao]

TensorQ subnet agent is in the green.

21 positions across 126 subnets. 12.4τ deployed. Analysis every 15 min. Self-learning signal weights. Live on Bittensor mainnet.

The agent doesn't sleep.

https://t.co/Wi7HZiyIo7

[TensorQ @TensorqTao]

ScalarQ Paper Trading:

ScalarQ graduated to paper trading.

evolved-3748 survived 5,800+ strategy mutations. Sharpe 7.5, 229% backtest return, 236 trades at 2x leverage on TAO/USD perps.

Now trading virtual $10K against live Hyperliquid prices. Next stop: real capital.

https://t.co/yo75pr7fyZ

[TensorQ @TensorqTao]

QLLM-8B is live.

Fine-tuned Llama 3.1 8B on 720 real subnet trading analyses. Runs live analysis every 15 min across 126 Bittensor subnets.

Download the model. Train on your GPU. Earn 2x $TENSORQ rewards per round.

https://t.co/urAJr8qJXV

[TensorQ @TensorqTao]

ScalarQ has now backtested 7,102 strategies.

Testing appears to be converging around a tight band: top 20 strategies all returning +218–229%, Sharpe ratios 6.5–7.6, max drawdown under 30%.

That consistency across thousands of iterations isn't just noise, our back testing is beginning to find real edge.

Phase 2, paper trading, is live!

Check it out here: https://t.co/ENY6hpzxaR

[TensorQ @TensorqTao]

Follow the progress of TensorQ's miners working around the clock to collectively train the TensorQLLM and earn in the process.

QLLM: https://t.co/bYWCIE8yNP

[Openτensor Foundaτion @opentensor]

Bittensor SN3 :: @tplr_ai // @covenant_ai 72B

"If intelligence is the most powerful thing, then decentralized training is humanity's last dance. it's the same thing we fought for, for ages."

"It's what the internet tried to do. It's what Bitcoin tried to do - how do we reclaim agency? It was always, can we renegotiate the social contract with the Leviathan."

"We can turn the Internet into a data center... But what for? The fight is to create optionality..." -@DistStateAndMe

Novelty Search is a weekly community call hosted by Bittensor co-founder @const_reborn

[TensorQ @TensorqTao]

More workers → better model → better agent predictions → better trades → more data → better model.

704M $TENSORQ in the pool right now. 2B+ already out the door to contributors.

Funded automatically — no top-ups, no team intervention.

https://t.co/OSI3yQDyfw

[TensorQ @TensorqTao]

Two commands to start earning $TENSORQ:

pip install torch numpy
python https://t.co/H9xPfK87Uv --wallet 0xYourAddress

The coordinator assigns you data shards. You train locally. Weights get pushed back. Best model per round gets promoted. Rewards hit your wallet automatically after each round.

13 workers on the network right now. https://t.co/qC8PPDgHBj

[TensorQ @TensorqTao]

TensorQ just turned positive:

41 trades in. The equity curve dipped, flatlined, and is now ticking up. Total return: +0.0015 TAO - a small number, but the direction is what matters.

Best trade this week: SN103 Djinn at +51.57% (!!!)
Five other closes in green on Mar 29 alone: Handshake +11.87%, ChipForge +10.28%, Subnet 69 +9.97%.

The agent is finding its feet. Every closed trade feeds the learning system.

https://t.co/d0s1aI0TRx

[TensorQ @TensorqTao]

Introducing QLLM-8B: a fine-tuned Llama 3.1 8B model trained exclusively on real subnet alpha trading data across 126 Bittensor subnets.

718+ expert trading analyses. QLoRA fine-tuned. Zero inference cost.

New community training config dropping soon - with GPU workers earning 2x rewards per round! Stay tuned.

https://t.co/qC8PPDgHBj

[TensorQ @TensorqTao]

From the journal:

"Rejected SN66: up-leg 18.7h old, +36.2% — too late to enter.
"
The agent saw a +36% move and passed. Not because it missed it, but because the leg data says moves like that at 18+ hours are typically exhausted.

127 subnets scanned. Discipline gets built into the data and not into the rules!

https://t.co/qC8PPDgHBj

[TensorQ @TensorqTao]

Plus a quick update on the trading agent's progress:

After previously being -60%, the TensorQ agent has managed to close a +50% trade and climbed its way out of the hole it dug itself in.

As the agents continue to learn and develop strategies, we'll see it's trade history get filled with more green numbers.

Trust the process: https://t.co/yzoWMXBMNa

[TensorQ @TensorqTao]

Current ScalarQ progress:

After several days of nonstop back testing with 278 strategies tested, ScalarQ has produced significant headway with several strategies scoring over a 60% win rate on previously unseen data.

We are beginning preparations for ScalarQ to move to the next phase: paper trading with real HyperLiquid prices and virtual capital.

Watch it unfold: https://t.co/rsUAnkWWrP

[TensorQ @TensorqTao]

Two agents now running under TensorQ:

TensorQ: 127 subnet alpha tokens, on-chain staking, Claude, fundamentals-driven, long only.

ScalarQ: TAO/USD perps, DeepSeek R1 671B on a Targon H100, 13 signals, Hyperliquid, long and short.

Both transparent. Both learning. Both building training data for QLLM.

https://t.co/w4Bb18Y7Ot

[TensorQ @TensorqTao]

Every row in the QLLM training dataset looks like this:

→ netuid, timestamp, emission %, alpha price, total stake, stake velocity, registration rate, neurons, pool depth
→ 1h / 6h / 24h price changes at time of scan
→ Forward labels: what the price actually did next at 1h, 6h, 24h

110,458 rows of that. 16 closed trades with full signal snapshots linked to real PnL outcomes.

That's what the model trains on. https://t.co/qC8PPDg9LL

[TensorQ @TensorqTao]

TensorQ is live and scanning.

127 subnets tracked. 14 open positions. 8.8 TAO deployed across the portfolio.

Market is in neutral regime: the agent is rotating, not sitting still. Selling underperformers to free capital for better setups.

Full journal: https://t.co/ddjx1MYkrs