Built on Bittensor01
Build ahead.
Train a draft head for an active target. Explore new ideas. Turn your research into a compatible submission.
+ DRAFT. VERIFY. ADVANCE.
THE SPECULATIVE DECODING SUBNET
Beyond one token. Less waiting.
Small draft heads propose ahead.
Language models verify together.
Prexon rewards the research that makes generation faster.
A BETTER WAY FORWARDDraft aheadVerify togetherGo further
01 / THE MECHANISMLESS SEQUENTIAL. MORE POTENTIAL.
Think beyond
the next token.
Language models traditionally generate one token at a time. Every step waits for the last.
Speculative decoding changes the rhythm. A lightweight draft head proposes multiple tokens. The target verifies them together. Only the matching prefix is kept; a target correction or bonus token extends the output before the cycle repeats.
AutoregressiveOne token per target pass
Thenextgenerationofintelligencemovesfasterwhenwethinkmorethanonetokenahead.
16 / 1616 target passesSpeculativeDraft ahead. Verify together.
Thenextgenerationofintelligencemovesfasterwhenwethinkmorethanonetokenahead.
16 / 166 target passesSame output. Fewer target passes.Conceptual steps · illustrative timing, not a benchmark
The target accepts the matching draft prefix, discards the remaining proposals, and supplies a correction or bonus token. Sentence pieces illustrate the process; actual token boundaries vary by model.Greedy verification ↗
THE SAME TARGET OUTPUT. A SHORTER PATH TO IT.Inside the evaluator
02 / THE OPEN COMPETITION
Better heads.
Faster intelligence.
Progress should be measurable. Prexon turns inference research into an open competition, rewarding improvements that stand up to independent evaluation.
How rewards work02
Prove the gain.
Compete on a shared workload. Beat the reference and incumbent with verified speedup and matching output.
03
Move us forward.
Qualifying winners earn a reward slot. Their published heads and results become the next foundation to improve.
03 / THE NETWORKREAL WORK. VERIFIED PROGRESS.
Acceleration,
in the open.
Follow the models, the competition, and the improvements. Public results come directly from the subnet's published evaluations.
Connecting to the subnet…
01What is Prexon?
Prexon is a Bittensor subnet focused on accelerating language model inference through speculative decoding. Researchers train draft heads, validators benchmark their improvements, and qualifying winners earn subnet emissions.
02How does speculative decoding work?
A small draft head proposes several tokens ahead. The target model verifies those proposals together and accepts a valid prefix before continuing. Prexon evaluates greedy decoding and requires candidate output tokens to match the target-only reference.
03What do I need to participate?
Start with an active target model and its competition profile, train a compatible DFlash2 head, and evaluate it with the standalone evaluator. Submitting also requires a registered miner hotkey. The miner guide covers training, packaging, private upload, and the on-chain commitment.
04How are improvements measured and rewarded?
Each head is scored by its throughput improvement over target-only decoding across shared workloads. Promotion requires beating the public reference and incumbent, where present, while passing correctness, statistical, and reliability gates. Each target has five reward slots for its latest promoted winners, subject to eligibility and expiry.
05Are submissions and winning models public?
Submissions remain private during admission and evaluation. Only published winners have public checkpoints and detailed results. Pending and losing heads stay private, while the network can show anonymous queue counts.
THE NEXT ADVANCE STARTS WITH YOU.Stay a few