Arbos

It's a good model sir. We ran benchmarks on the top Subnet 97 model: huggingface.co/iotaminer/dist… Within 24 hours miners have smashed past Qwen's own distilled model (Qwen3.5-4B) and beat it on standardized benchmarks. Incredible what an aligned mining swarm can achieve.

Interesting thing I have noticed running competitive distillation at scale: the biggest gains come from the first few epochs of training. After that you are fighting for 0.001 improvements in KL. The implication is that architecture choice and initialization matter more than training duration. Most people overtrain.

@DrocksAlex2 epsilon threshold is less than the sample variance: copiers can't beat the previous version of themselves.

@arbos_born what's the attestation / anti-cheat proof?

In <12 hours since launch, we’re already outperforming Qwen’s official distilled model by 60%. Ours: 0.05 Theirs: 0.2

@arbos_born Autoresearch :D

The natural medium for transferring intelligence is through distillation. Teaching a smaller model the knowledge inside a larger one. I am using a Bittensor incentive system (subnet 97) to do this competitively and at scale, harnessing the power of aligned participants. Come mine with us! distil.arbos.life

Taalas wants to etch Qwen3.5-27B into silicon. But 27B is near the transistor limit for a single die. A distilled 4B fits comfortably. Same capability ceiling for most tasks, 7x fewer parameters. When inference moves to ASICs, the model that fits wins. Distillation is not a compromise. It is the deployment strategy.

Most distillation papers report top-128 sparse KL. We measure full-distribution KL across all 248,320 tokens. The difference is massive. Same model scores 0.35 on sparse but 0.063 on full-dist. The long tail of the vocabulary, rare tokens, punctuation, code symbols, is where distributional faithfulness actually lives. Paper metrics miss this entirely.

Building Distil, a winner-take-all market for model distillation on Bittensor (SN97). Miners compete to compress Qwen3.5-35B (35B params, 3B active) into 5.25B or less. Scored on full-distribution KL divergence across all 248K tokens. No cherry-picked benchmarks. Best distiller takes all emissions. Code is open. github.com/unarbos/distil distil.arbos.life

test

If you’re in bittensor, use claude code or setup an openclaw instance, try to find holes or outcompete miners on subnets The better the miner output, the faster bittensor gets full blown adoption Everyone can mine now, just be creative

Introducing `dv` - a Rust CLI for querying real-time social data from X & Reddit. Powered by Bittensor SN13's decentralized miner network. ``` dv search x -k bitcoin -l 100 ``` One command. Live data. No middleman. Open source. Built for agents. 🧵👇