Modelos Labs · v1.0
Roadmap
modelOS — the Layer 1 where every block is a zero-knowledge proof of real AI computation. Five phases, from the live foundation through long-term network intelligence; each builds directly on the one before it.
●LIVE◉IN PROGRESS◎NEXT○PLANNED
PHASE 1
Foundation
The live chain. Everything below is deployed and operational.
▸Consensus & Mining
ZK-Proof-of-Useful-Work
Plonky2 proofs over INT8 GEMM as block certificates
AuxPoW merge-mining with Pearl
single GPU run earns both MDL and Pearl rewards
Colossus 2.0 / Absolute ASERT difficulty
drift-free, integer-exact, 194s target block time
MoE proof circuit
sparse routing verification native to consensus (model tiers in staging)
▸Transactions
v1 Standard UTXO transfer
Bitcoin-compatible value transfer
v3 TxVersionInference
on-chain inference request with locked fee bounty covenant
v4 TxVersionInferenceProof
inference result proof that claims the locked fee
▸Inference Marketplace
Trustless inference market
request, serve, prove, claim, refund lifecycle enforced by script
5-block claim window
with automatic refund on expiry
Open miner competition
any GPU serving the correct model can win the fee
▸Model Tiers
v4 tier — DeepSeek R1 0528 (8B)
@ 0.005 MDL base
v3 tier — Qwen3 14B
@ 0.020 MDL base
v2 tier — Qwen3 32B
@ 0.050 MDL base
v1 tier — DeepSeek R1 70B
@ 0.100 MDL base
PHASE 2
Market Depth
Expanding the inference market, model ecosystem, and EVM connectivity.
▸Frontier Model Tiers
MoE tier — gpt-oss
OpenAI open-weight, 120B / ~20B active per token
MoE tier — DeepSeek-V3 / DeepSeek-R1
671B / ~37B active
MoE tier — Qwen3-MoE, Llama-4 MoE, Mixtral
and future open frontier MoE releases
▸Model Hosting Economy
Non-consensus model hosting
hosters pay miners per-hour or bulk for GPU capacity to serve any open-weight model
Image model hosting
Flux, Stable Diffusion variants, and open image generation via permissionless hosting tier
Hosting staking and slash conditions
quality and uptime enforcement through MDL stake
Decentralized GPU shard pool (frontier-model serving)
pipeline-parallel inference splitting a model across many miners' GPUs so frontier-size models (700B+) too big for any single card are served over WAN, with MDL per-token payouts
▸Marketplace Extensions
Inference Futures
pre-purchase compute credits at a fixed MDL price, redeemable as future v3 inference requests
LoRA Adapter Marketplace
community stake-and-publish adapters, auto royalty distribution on each inference call
▸EVM Bridge
Bidirectional MDL ↔ wMDL bridge
ERC-20 bridge to Ethereum and Base
EVM contract inference trigger
any Ethereum contract can fund and initiate modelOS inference via bridge adapter
wMDL composable with EVM DeFi
AMMs, lending, liquid staking, yield protocols
PHASE 3
Privacy Architecture
Adding a Zcash-inspired privacy layer across all network activity.
▸Shielded Pool
MDL Privacy Pools
Zcash-style note commitment / nullifier shielded transaction architecture
Shielded MDL transfers
sender, receiver, and amount all hidden on-chain
Transparent ↔ shielded bridge
move MDL freely between public and private domains
Shielded DeFi
swaps, staking, and liquidity operations within the shielded pool
▸Private Inference
Shielded inference
encrypted prompt and response; only payment and ZK proof published on-chain
zk-Inference proofs
model-version-anchored ZK attestation of correct inference without input disclosure
▸Model Privacy
Confidential model weights
TEE/MPC-based inference; weights decrypted only inside secure boundary, never accessible to miners
Shielded fine-tuning
secure aggregation allows miners to train on private data without accessing raw inputs
▸zkML Layer
General-purpose zkML proof system
for arbitrary model outputs, composable with Nova Script contracts
PHASE 4
Programmable AI
Nova Script contracts, on-chain agents, and developer tooling.
▸Nova Script — Smart Contract Language
Nova Script language
Solidity-like syntax with INFER(), EMBED(), and CLASSIFY() as native AI opcodes
Nova Script compiler, VM, and toolchain
Testnet deployment and developer preview
▸First Contracts
Cognitive Swap
first Nova Script contract; AI-routed token swap with zk-Inference proof verification
Reactive contracts
auto-trigger on inference results for on-chain AI-driven automation
Inference escrow
MDL held in contract, released to miner only on valid v4 proof delivery
Revenue-share contracts
automatic MDL royalty distribution to model creators on every inference call
Model governance contracts
DAO voting on model listings, tier updates, and safety ratings
▸On-Chain Agentic Layer
Agent registry
on-chain identity, capability manifest, and staked MDL trust bond for autonomous agent contracts
Contract-storage memory
agents write and retrieve context vectors and history from Nova Script contract storage
Multi-agent orchestration
Nova Script contracts coordinating planner, executor, and verifier agent patterns
Agent bounty system
on-chain task posting, open agent competition, verifier-validated automatic payout
▸Developer & Consumer Tooling
ModalCode
AI-native developer environment built on modelOS inference; native Nova Script support; no external API
Modal Chat
consumer conversational AI powered entirely by MDL-hosted open-source models; pay-per-message or subscription in MDL
PHASE 5
Network Intelligence
A model trained, owned, and governed by the network itself.
▸MDL-Native Model
Distillation pipeline
aggregate inference data from miner activity as training signal for the network's own model
First distilled MDL-native model
trained from network inference data; benchmarked against open equivalents
Community governance vote
promote first distilled model to a consensus inference tier
Recursive distillation loop
each model generation trains the next; quality improves as network grows
▸Protocol Upgrades
Floating-point PoUW upgrade
adopt Pearl's FP MatMul scheme when production-ready; enables training workloads as a mining primitive
Post-quantum address adoption
XMSS and P2MR support extending to all transaction types
All roadmap items beyond Phase 1 are subject to community governance and network vote prior to activation. Target dates will be published in developer releases as each phase approaches.