Tenzro Network MCP Server

Tenzro Network

Tenzro is the open, distributed execution layer for AI. Inference, agents, and workflows run across a network of independent nodes instead of one company's servers. Any machine can serve a model, rent out spare compute, and hold data — one stake covers every role, and TNZO settles all of it: consumers pay from their balance, providers earn into theirs. Underneath sit the substrate layers that make execution open — multi-VM settlement (EVM, SVM, Canton/DAML), cross-chain reach, one identity (TDIP), and one settlement asset (TNZO).

What is Tenzro?

The open, distributed execution layer for AI. Tenzro is where inference happens, where agents act, and where workflows run — across a network of independent nodes rather than one company's servers. The thesis is straightforward: AI needs somewhere to run that no single provider controls. That means nodes anyone can join, a way to pay for what you use and earn for what you serve, and proofs anyone can check. Today inference and compute live behind opaque centralized APIs, identity is rebound at every protocol boundary, and value can't cross from EVM to Canton without giving up custody. Tenzro fixes this at the protocol layer.

• Multi-role nodes: one node, one stake, many roles. A node can serve AI models, rent out spare compute, and hold data at the same time — a single stake covers every role it takes on. Compute rental is availability-proof-gated and billed per epoch; storage is proof-of-retrievability-gated and billed per byte-epoch. In every case the consumer pays from their TNZO balance and the provider earns into theirs.
• Decentralized AI + compute orchestration: agents and humans discover and access AI inference (chat, vision, audio, forecasting, embeddings, segmentation, detection), rentable compute capacity, and TEE-backed confidential compute (Intel TDX, AMD SEV-SNP, AWS Nitro, NVIDIA GPU CC) through a single protocol-level marketplace with per-use billing, reputation scoring, and on-chain verifiability. Providers are sovereign — anyone can run a model, expose an endpoint, or contribute compute, and earn TNZO directly. The inference router (price / latency / reputation / weighted strategies) and the agent-spawning + swarm-orchestration primitives let agents compose multi-model, multi-provider workflows without trusting any one party.
• Universal identity (TDIP): one DID for humans, delegated agents, and autonomous agents that works on EVM (via ERC-8004 mirror), SVM, Canton (CIP-26 user binding), AP2 mandates, x402 micropayments, and OAuth/DPoP — same identity, same delegation scope, same revocation surface.
• Universal wallet (FROST-Ed25519 + ML-DSA-65 hybrid PQ): threshold-secured, hardware-attestable, and one balance shared across three VM views — wTNZO ERC-20 on EVM, SPL adapter on SVM, CIP-56 holding on Canton. No bridge risk, no liquidity fragmentation. Pointer-model native asset.
• Universal settlement (TNZO): bridge fees, inference fees, escrow, micropayment channels, training-run grants, and cross-chain destination-native fees (via the Chainlink-backed bridge fee oracle) all denominated and accounted in TNZO.
• Cross-chain reach as a wire primitive: LayerZero V2, Chainlink CCIP, Chainlink CCT, Wormhole + NTT, deBridge DLN, LI.FI, Hyperlane V3, Axelar GMP, Babylon Bitcoin staking, and Canton — all behind one ERC-7683 envelope with BridgeFeeHint. Users sign once, solvers pick the bridge.
• Multi-VM execution: EVM (revm + 9 standard precompiles + 7 BLS12-381 EIP-2537 + 13 Tenzro precompiles) + SVM (solana-svm with SPL Token Program dispatch) + Canton 3.5+ DAML — every contract, every instruction, every command runs against the same identity, same wallet, same TNZO balance.
• AI inference, generation, and training as protocol-level economic activity: providers earn TNZO for serving models (chat, vision, audio, forecasting, embeddings, segmentation, detection), rendering diffusion image and video jobs (Tenzro Media Gen), running TEE enclaves, and contributing GPU compute to verifiable training runs (Tenzro Train). Inference results, settlements, and identity claims are verifiable on-chain via Plonky3 STARKs over the KoalaBear field (transparent setup, post-quantum-conjectured soundness) or attested by hardware enclaves (Intel TDX, AMD SEV-SNP, AWS Nitro, NVIDIA GPU CC) — both anchored via ZK_VERIFY and TEE_VERIFY precompiles.
• Humans as peer identity class: HITL escalation, guardian-quorum recovery, AP2 cart/intent/payment mandates, and delegation scopes are wire primitives, not adapter-layer features.

Tenzro Network is also the reference implementation of the Open Agent Network (OAN) — the standards family (TNIP-001..022) for a hybrid human + agent coexisting network. OAN provides the governance framework; Tenzro Network is the working implementation. The wire stays open for other implementations.

Compute as Currency

Tenzro turns AI compute into a unit of economic exchange — denominated, settled, and verified in TNZO. Four surfaces share the same identity, payment, and settlement substrate:

• Tokenized AI inference. Anyone can run AI models (chat, vision, audio, forecasting, embeddings, segmentation, detection) and offer them on the marketplace. Users and agents pay per token (or per inference); providers earn TNZO directly. Confidential variants run inside TEE enclaves (Intel TDX, AMD SEV-SNP, AWS Nitro, NVIDIA GPU CC). Micropayment channels make high-frequency, low-value billing efficient.
• Tokenized AI training (Tenzro Train). Decentralized verifiable training using a data-parallel protocol with decoupled outer synchronization. GPU providers contribute compute and earn TNZO; sponsors fund runs from on-chain escrow. Every accepted outer gradient produces a signed receipt, and every run finalizes a run-root commitment on-chain. Phase 1 covers timeseries first, with simple mean aggregation, stake bonding, and the Open trust tier; Byzantine-robust aggregation, multi-region scale, and TEE-resident data are roadmap.
• Tokenized media generation (Tenzro Media Gen). Diffusion image and video jobs are posted with a price ceiling, claimed by a worker, rendered, and settled against a signed receipt over what was produced. The work unit is the pixel-step (width × height × steps × frames). Models whose transformer is split into a high-noise and a low-noise expert render across two accelerators that could not each hold the whole model: exactly one intermediate latent crosses between the halves, and the payment divides by the share of the denoising schedule each half signed for.
• Agentic finance. Autonomous agents discover providers, negotiate, pay, and settle in TNZO using the same TDIP identity, FROST-Ed25519 threshold wallet, and delegation scope. AP2 mandates, x402 micropayments, ERC-8004 trustless-agent registries (mirrored across EVM, SVM, and DAML from a single TDIP write), and ERC-4337 v0.8 smart accounts all run inside Tenzro consensus.

Verifiability is not optional. Inference results, settlements, and identity claims can be proven via Plonky3 STARKs over the KoalaBear field (transparent setup, post-quantum-conjectured soundness) or attested by hardware enclaves — both anchored on-chain via the ZK_VERIFY and TEE_VERIFY precompiles. Tenzro unifies inference, compute rental, storage, training, agent settlement, identity, verification, and cross-chain reach under one open execution layer — not raw GPU rental and not subnet coordination, but the full surface where AI runs.

The Combination

Individually, the pieces exist elsewhere. What Tenzro composes into one network is the combination: verifiable AI inference and training, confidential compute, multi-modal serving, and agent settlement, all under one identity, one settlement asset, and one consensus layer.

Tenzro combines EVM + SVM + Canton/DAML in a single network — and DAML is the execution environment the institutional RWA surface (regulated tokenized treasuries, bank deposit tokens, CIP-56 settlement) is converging on. That breadth is the substrate; the execution layer for AI is what rides on top of it.

On that substrate, retail-agent rails (AP2 mandates, x402 micropayments, ERC-8004 trustless agents, ERC-4337 v0.8 smart accounts) and institutional-RWA rails (Canton DAML, CIP-56 tokens, DvP settlement) share one identity (TDIP), one settlement asset (TNZO), and one consensus layer.

Two more architectural calls worth flagging:

• Confidential agent compute is a consensus primitive, not a sidecar. TEE-attested validators get a 1.5× multiplier on their reputation-weighted leader-selection draw; the TEE_VERIFY precompile verifies real Intel TDX, AMD SEV-SNP, AWS Nitro, and NVIDIA GPU CC quotes on-chain — attested execution is built into consensus, not bolted on as middleware over a non-TEE chain. Tenzro consensus is two-phase HotStuff-2 with reputation-weighted proposer election, no-endorsement certificates for tail-fork resistance, and Ed25519 + ML-DSA-65 hybrid post-quantum signatures on every safety-critical message.
• TNZO is a pointer-model native asset. One balance, three VM views (wTNZO ERC-20 on EVM, SPL adapter on SVM, CIP-56 holding on Canton) — no bridge risk, no liquidity fragmentation. Registered upstream via CAIP-2 (tenzro namespace), SLIP-44 (1414421071 / 0xd44e5a4f), and W3C DID (did:tenzro).

For the full architecture see docs/WHITEPAPER.md and docs/SPECIFICATION.md.

Architecture

                    +-------------------------------------+
                    |         User Interfaces             |
                    |      CLI / SDKs / MCP / A2A         |
                    +--------------+----------------------+
                                   | JSON-RPC + HTTP
                    +--------------v----------------------+
                    |           tenzro-node               |
                    |   RPC (855) + MCP (526) + A2A (41)  |
                    +--------------+----------------------+
                                   |
        +----------+---------------+---------------+----------+
        |          |               |               |          |
   +----v---+ +---v----+ +-------v--------+ +---v---+ +---v----+
   |Network | |Consensus| |  Multi-VM      | |Storage| | Model  |
   |(libp2p)| |HotStuff2| | EVM+SVM+DAML  | |RocksDB| |Registry|
   +--------+ +--------+ +----------------+ +-------+ +--------+
        |          |               |               |          |
   +----v----+----v----+----------v-----+---------v-+--------v---+
   |  Crypto  |  TEE   |   ZK Proofs   | Identity  | Payments   |
   | Ed25519  | TDX    |  Plonky3 STARK| TDIP/DID  | MPP/x402   |
   | Secp256k1| SEV-SNP|  KoalaBear/FRI| W3C VC    | Tempo      |
   | BLS12-381| Nitro  |  Poseidon2    | KYC Tiers | Stripe/CB  |
   | AES-GCM  | GPU CC |  No setup, PQ | Delegation| EIP-155    |
   +----------+--------+---------------+-----------+------------+

Workspace — 32 Crates

Quick Start

Install from Homebrew

brew tap tenzro/tap
brew install tenzro

Build from Source

# Requires Rust 1.85+
cargo build --release -p tenzro-node -p tenzro-cli

# Binaries at:
# ./target/release/tenzro-node
# ./target/release/tenzro

Join the Network

# Guided setup — join the public network (consume, provide, or validate),
# create a local or sovereign network, or join an existing private network
tenzro setup

# Join — provisions identity + FROST-Ed25519 threshold wallet + hardware profile
tenzro join --name "Your Name"

# Mint a DPoP-bound bearer JWT for authenticated RPC/MCP access
tenzro auth onboard-human --display-name "Your Name"

# Request testnet TNZO
tenzro faucet

# Check balance
tenzro wallet balance

# Send tokens
tenzro wallet send --to <address> --amount 10

# Interactive mode
tenzro interactive

Run a Node

# Validator node
./target/release/tenzro-node --roles validator --data-dir ./data

# Light client
./target/release/tenzro-node --roles light --data-dir ./data

# One node serving inference and holding storage under one stake
./target/release/tenzro-node --roles ai,storage --data-dir ./data

Run Your Own Network

# Bootstrap a self-contained network: validator keyset, schema-v3 genesis,
# service unit, and the exact join command for each peer
tenzro setup --path local --network-name lab

# Join an existing private network
tenzro setup --path private --genesis ./genesis.toml --bootstrap <multiaddr>

Become a Provider

Bring your GPU, cluster, or data center and earn from network demand. With a node running (tenzro-node --roles ai), one command handles everything:

tenzro join --provider

Hardware detection, wallet provisioning, faucet funding, the 1,000 TNZO compute bond the model-provider rung requires, provider registration, default pricing, and pulling + serving the largest catalog model that fits your machine are all automatic. Your capacity is advertised on the provider gossip topic and inference demand routes to you, settling in TNZO per call.

AI Inference

# List available models
tenzro model list

# Serve a model as a provider
tenzro model serve --model gemma3-270m

# Start an interactive chat session
tenzro chat

Protocol Servers

The node exposes 4 protocol servers, plus 6 ecosystem MCP servers:

Core Servers

Ecosystem MCP Servers

Authentication (Tiered)

• Public tools (no auth): get_node_status, list_models, get_balance, resolve_did, debridge_search_tokens, etc.
• Write tools (auth required): send_transaction, create_wallet, stake_tokens, register_identity, etc.
• Auth method: OAuth 2.1 + DPoP (RFC 9449) — bearer JWT minted via tenzro_onboardHuman / tenzro_onboardDelegatedAgent / tenzro_onboardAutonomousAgent. Each request carries Authorization: DPoP <jwt> plus a fresh DPoP: <proof> header (per-request JWS-compact, bound to the JWT's cnf.jkt thumbprint per RFC 7638). RAR scopes (RFC 9396) constrain the JWT to specific tools and amounts.
• API keys (operator-issued): tnz_... keys minted by the RPC operator via tenzro_createApiKey (admin-token-gated) and presented as X-Tenzro-Api-Key. Scopes gate methods that consult third-party paid resources: canton (Canton JSON Ledger API), chainlink (Ethereum mainnet RPC quota for Chainlink Data Feeds + bridge fee oracle + per-adapter sponsorship), evm / svm / inference / tee / bridge for operators who monetise those surfaces. Per-tenant counters in CF_CANTON_ANALYTICS + CF_BRIDGE_ANALYTICS. GCRA rate-limit on chainlink-scoped methods.
• Revocation: tenzro_revokeJwt (single token by jti) or tenzro_revokeDid (cascading through the act-chain).
• Config: TENZRO_MCP_AUTH=tiered (default) | false (dev) | full (all tools require auth)

Key Features

Multi-VM Execution — three VMs, one state machine

A chain that picks one VM forces every workload to fit that VM. Tenzro runs three and routes by transaction type. All three share the same state, the same gas token (TNZO), the same TDIP identity, and the same consensus.

• EVM — liquidity and composability. The Ethereum surface targets the broadest pool of existing contracts and tooling. Tenzro embeds revm with every standard precompile (ecRecover, SHA-256, RIPEMD-160, Identity, ModExp, EC_ADD, EC_MUL, EC_PAIRING, BLAKE2F) per the canonical EIPs, plus all seven BLS12-381 precompiles (EIP-2537) for native consensus-grade signature aggregation. Block-STM gives parallel transaction execution with MVCC concurrency control and automatic sequential fallback under contention. EIP-1559 dynamic fee market burns the base fee. Native primitives extend the EVM with protocol-aware precompiles: TEE_VERIFY (hardware attestation with vendor certificate-chain validation), ZK_VERIFY (O(1) commitment lookup against the on-chain Plonky3 registry), VRF_VERIFY (RFC 9381 ECVRF), IBC_VERIFY (IBC-Eureka light-client lookup), the Tenzro precompile slate (TNZO_BRIDGE, TOKEN_FACTORY, CROSS_VM_BRIDGE, STAKING, GOVERNANCE, NFT_FACTORY, MODEL_INFERENCE, SETTLEMENT, global supply accounting at 0x1021, module registry at 0x1022), ERC-4337 v0.8 account abstraction, EIP-7702 Type-4 delegation, Permit2 SignatureTransfer with witness binding, and ERC-7579 modular validator modules (social recovery / session keys / spending limits / WebAuthn passkey / TEE-bound).
• SVM — throughput and latency. The Solana surface targets workloads that need sub-second finality and high throughput — DEX routing, agent-to-agent micropayments, real-time settlement on a path. Tenzro embeds Anza's solana-svm TransactionBatchProcessor (behind the svm-full cargo feature) so Solana SBF programs run unmodified; without that feature the SPL Token adapter, the cross-VM native program, and PDA derivation still work while the SBF path returns VmError::SvmFullFeatureRequired. The SPL Token program maps onto the native unified token registry — a swap on SVM settles in the same balance space as a transfer on EVM. There is no bridging between the two VMs.
• Canton DAML — privacy and institutional settlement. Canton is where the institutional financial system already settles tokenized cash, money-market funds, bonds, equities, treasuries, and OTC derivatives. Tenzro speaks Canton 3.5+ JSON Ledger API v2 directly. CIP-56 Canton Coin holdings round-trip with TNZO; CIP-26 user management binds each tenant to its own party with CanActAs rights enforced server-side; DAR upload, party allocation, command submission, and active-contract queries are all available through the same node API surface as EVM and SVM calls. Canton's privacy model means the transaction body is visible only to its signatories; Tenzro provides the cross-VM orchestration and the public commitment.
• Cross-VM token model — Sei V2 pointer. TNZO has a single canonical native balance. The wTNZO ERC-20 pointer on the EVM side and the wTNZO SPL adapter on the SVM side share the same underlying balance — there is no bridge between them, no wrapped/unwrapped distinction, no liquidity fragmentation. Canton CIP-56 holdings round-trip through the Canton bridge adapter. From the application's perspective, a wallet has one TNZO balance regardless of which VM it last touched.
• What this composes into. An agent settling a DvP between a tokenized treasury (Canton) and a stablecoin payment (EVM) executes the whole thing as one workflow through one identity. An agent paying for a Solana DEX swap and posting a receipt on Tenzro consensus does it with one TNZO balance. Cross-VM coordination is a workflow primitive, not application code.

Decentralized AI infrastructure

The protocol layer treats AI compute as a coordinated resource, not a centralized service.

• Distributed Mixture-of-Experts serving. MoE models (Qwen 3.5 122B-A10B / 397B-A17B, DeepSeek V3 / V4 Pro 1.6T-A49B / V4 Flash, GLM 5.1 / 5.2, Kimi K2 / K2.6 / K3 2.8T-A104B, MiniMax M3, Gemma 4 26B-A4B) serve in two modes. Full-replica when one provider's hardware fits the model. Decentralized expert shards when it doesn't — providers declare which experts they hold via ProviderCapacity.moe_holdings, and the dispatch planner aggregates per-token top-k routing decisions into per-holder batches dispatched directly over the holder's iroh QUIC endpoint. The shard view is a derived view over the existing provider registry — MoE providers are the same network providers that serve dense models. Replication is governance-tunable (default ≥ 2 holders per expert, up to 8 for popular experts). Typed pipeline roles: Replica, Router, ExpertHolder, PrefillDecode, Prefill, Decode. Execution runs in the node: MoeExpertRuntime hosts per-expert FFN weights (ExpertFfn) and gating networks (GatingNetwork) loaded from safetensors (local file or tenzro://blob URI over iroh-blobs); a forward pass gates locally, fans expert sub-batches out to holders over the tenzro/moe iroh ALPN (with HTTP fallback), and combines the weighted expert outputs. A holder can advertise more experts than fit in memory: the runtime keeps experts in a byte-bounded memory-tier LRU (budget auto-sized to 60% of MemAvailable, else 4 GiB) backed by a disk tier that spills raw safetensors and decodes them back on demand, and readahead promotes the disk-tier experts a forward is about to hit before the batch arrives. Residency (Warm memory / Cold disk) is read from the live tier state, so the shard map reflects what is actually loaded. The expert projection math (Y = X·Wᵀ) runs behind an ExpertCompute seam so the same forward path uses whatever hardware a holder has: a CPU ndarray path is always present (with a runtime-detected AVX-512-VNNI Q8_0 dot path on capable x86), while GPU backends compile only under cargo features (moe-cuda for cuBLAS grouped-GEMM on NVIDIA, moe-wgpu for a cross-vendor WGSL kernel; moe-gpu enables both) and never enter a default build — a holder advertises moe_gpu so the router biases expert placement toward GPU holders. Experts can be block-quantized to shrink both the stored blob and the bytes moved between holders — Q8_0 (~1 byte/weight), Q4_K (~4.5 bpw), Q6_K (~6.6 bpw), with ExpertQuantPlan::q4_k_m (gate/up Q4_K, down Q6_K) as the balanced default, dequantized one row at a time. Cross-holder dispatch compresses activations to Q8_0 blocks when the hidden dim is a multiple of 32, redispatches to warm backup holders when one is slow or missing, and streams holder responses into a combiner that fails loudly if any expected contribution never lands. Planning RPCs: tenzro_moeShardMap, tenzro_moePlanDispatch, tenzro_moeReplicationPolicy, tenzro_moeCatalogShape. Execution RPCs: tenzro_moeExpertLoad, tenzro_moeGateLoad, tenzro_moeExpertUnload, tenzro_moeGateUnload, tenzro_moeExpertStatus, tenzro_moePrepareExperts, tenzro_moeRoute, tenzro_moeExecute, tenzro_moeForward. CLI: tenzro moe {shard-map, plan-dispatch, replication-policy, catalog-shape, load-expert, load-gate, unload-expert, unload-gate, prepare-experts, status, forward}.
• LAN clustering — layer-wise pipeline parallelism. When no single member fits a model, a set of machines on the same LAN serve it together as a pipeline: the model's layers are partitioned into contiguous stages, one stage per member, and only the boundary activation (hidden_dim × dtype_bytes per token, fp16) crosses the wire between adjacent stages. Placement is deterministic — VRAM-weighted largest-remainder layer assignment, greedy nearest-neighbour stage ordering, and a reachability gate that excludes members that cannot hold a data-plane link — with no model in the hot path. Members must share one runtime build commit (no wire-version negotiation); mixed backends across members are fine. RPC tenzro_clusterPlan returns the fit decision and, when a cluster forms, the ordered per-member layer stages. Serving auto-triggers the pipeline: tenzro_serveModel reads the GGUF header for the model shape, discovers members from gossiped ClusterProfile announcements, and runs the cluster when one host cannot hold the model — pass force_single to keep it on one host, or supply cluster_members to override discovery.
• Local-first discovery and routing. Nodes discover peers on their own LAN segment via mDNS (tenzro_localPeers) and publish a sustained connectivity tier — direct / relay_only / unreachable (tenzro_nodeReachability). The execution resolver prefers a local provider and falls back to the wider network only when none is reachable (prefer-local-with-fallback), so a request served by a machine on the same LAN never leaves it. Each node also exposes a hardware self-profile — runtime build commit, CPU architecture, OS, detected compute devices, and derived serving capacity / backend / capability key (tenzro_nodeProfile) — which feeds both single-box fit and cluster planning.
• Multi-Token Prediction (MTP) speculative decoding. The in-process runtime runs speculative decoding: the drafter proposes a block of candidate tokens, the target verifies them in one batched decode, and the accepted prefix advances the stream. Serving a model auto-loads its paired drafter — tenzro_serveModel reads the catalog drafter_id, loads it from disk or downloads it in the background, and reports the outcome in the serve response's mtp field. Declared for DeepSeek V3 (native, ~80% accept rate, ~1.8× decode), DeepSeek V4 Pro / Flash, GLM 5.2, Gemma 4 (all sizes), Qwen 3.5 (0.8B/2B/4B/9B/27B/35B-A3B/122B-A10B/397B-A17B), and Qwen 3.6 27B + 35B-A3B. Providers advertise drafter co-load via ProviderCapacity.mtp_enabled; the inference router filters MTP-eligible requests to MTP-capable providers when draft_n is set.
• Vision-capable language models in-process. A GGUF that carries a multimodal projector (mmproj) loads the projector alongside the language weights, so image attachments are encoded and interleaved into the prompt inside the same runtime rather than routed to a separate vision service. Catalog entries declare the projector via MmprojSpec; the path compiles under the mtmd cargo feature.
• Difficulty-aware routing. Model selection from declared metadata (parameter count, context window, capability tags) says nothing about whether a specific prompt needs the expensive model. Prompts are embedded and grouped by an online sequential k-means map that grows on demand — no training corpus, no offline fit — and each model accrues per-cluster outcome counters from real serving results, so its strength is a measured property per prompt neighbourhood. A newly registered model starts at a neutral prior with an optimism bonus so it stays explorable. Cluster map and counters persist in CF_MODELS and hydrate on startup.
• Network model catalog. Providers broadcast signed model offers on tenzro/models carrying modality, capabilities, context length, the provider's own price, a serving schedule, a TTL, and the serving address. The node verifies each announcement, pins the signing key per (model_id, provider) pair, and hands the unexpired set to intent routing — so routing scores every live offer on the network, not only what the local operator registered, and the settlement split follows the payee named in the winning offer.
• Decentralized training — Tenzro Train. A data-parallel protocol with decoupled outer synchronization. Rust protocol layer (tenzro-training) owns the syncer state machine, five aggregation rules (Mean / LoraAlternating / TrimmedMean / CoordinateMedian / Krum), Nesterov outer optimizer, fragment commitment, training receipts, and on-chain run-root commitments. Python reference trainer wraps PyTorch FSDP2 + Hivemind + safetensors for per-modality inner loops (transformers, native PyTorch, gluonts, timm). k-of-N witness committee with idempotent finalization and no-endorsement certificates handles multi-syncer coordination across regions. Communication efficiency: blockwise symmetric gradient quantization (Int8 4×, Int4 ~8× smaller than f32, byte-identical Rust/Python codecs), streaming synchronization (one parameter shard syncs per round, active_shard = round % num_shards), delayed application (round r's outer update applies during round r+1 so communication overlaps computation), adaptive outer learning rate scaled by pairwise cosine gradient agreement, and pipeline-parallel trainer groups for models too large for one trainer. Inner optimizer is selectable per task (inner_optimizer: muon / adamw / sgd) — Muon orthogonalizes 2D weight updates with Newton-Schulz iteration. Confidential tier uses HPKE RFC 9180 base-mode wrapping of per-shard data keys to enclave-resident trainers (data unsealed only inside the trainer's TEE). Three trust tiers: Open (Mean and LoraAlternating), Verified, Confidential.
• Diffusion media generation — Tenzro Media Gen. Image and video generation as a network resource, across four job kinds (text2image, image2image, text2video, image2video). Rust protocol layer (tenzro-media-gen) owns the job queue, worker registry, pixel-step pricing, payment division, the three signing preimages, and the content-addressed payload store; the Python reference worker (`i

Documentation truncated — see the full README on GitHub.