LIMONATA COMPUTE — PROVIDER GUIDE This testnet release supports automatic registration and a trusted coordinator. Providers run AI inference, not block validation or model training. User deposits pay for accepted requests; simply connecting a GPU does not mint LIMO or earn a reward. CURRENT RELEASE — WINDOWS 0.8.0 / LINUX 0.8.0 Default contribution keeps an assigned stage of one shared Qwen3 32B model resident on each compatible GPU. Every required stage must be ready before that model can answer. More compatible contributors add stage capacity or warm replicas; idle replicas are not counted as having processed requests. There is no smaller-model substitution when the shared model is unavailable. Only the assigned stage downloads: about 14.99 GB (layers 0–47) or 5.86 GB (layers 48–63). Allow at least 18 GB or 8 GB free disk respectively. Initial free VRAM admission is 18,623 MiB or 8,379 MiB respectively, with reserve included; a real full-GPU stage probe must still pass before readiness. These readings are provider-reported, not hardware attestation. Context is 512 with two bounded slots per stage. First setup needs a large download and GPU load; later requests reuse the resident stage. Restarting verifies disk files and reloads VRAM. Interrupted stage downloads resume; complete SHA-256 verification remains mandatory. Shared-mode requests are currently sponsored and earn no provider LIMO. Only stages actually used in accepted requests receive work-participation records. Connected time, reported VRAM, local tests and unused replicas earn no reward. There is no promised performance or income. Network distance and GPU speed matter. A required-stage loss can interrupt a request if no compatible warm replica recovers. Linux / Runpod NVIDIA CUDA guide: https://limonata.xyz/ai/providers/linux After installation, run limonata-provider start with the same --data-dir; provide --payout YOUR_PUBLIC_ADDRESS only when creating the first identity. Windows: extract the complete ZIP, run Limonata-Provider.exe, enter the public payout address and click Start contributing. Keep the console open. Neither platform requires manual operator credentials or secret configuration. Registration and capacity assignment precede the stage download. Readiness is advertised only after full GPU offload and a real assigned-stage probe. Independent whole-model work remains an explicit mode: one GPU keeps a complete pinned Qwen 1.5B, 8B or 32B model and serves matching jobs. Existing paid independent requests may earn test LIMO via escrow. Legacy tensor RPC remains an explicit advanced mode with its existing weight cache. A tensor cache is not a stage file. No mode silently relabels smaller-model answers as the shared 32B model. WINDOWS AMD / NVIDIA Windows 10/11 x64 with a compatible Vulkan GPU and current driver. AMD does not need CUDA or ROCm. Runtime enumeration may need the Microsoft Visual C++ 2015–2022 x64 runtime. The app does not install GPU drivers or require administrator access, Node.js, Python or WSL. Windows GPU execution remains unvalidated in our Linux build environment. Use OS security protections normally; sources/checksums ship with this unsigned preview. CPU fallback is rejected. Saved Windows identity uses DPAPI at %LOCALAPPDATA%\Limonata\Provider-v0.2\provider-identity.dat. Upgrading with the same Windows account reuses identity, payout and old caches; do not delete the identity to retry. The provider generates its own random token, keeps it private, and sends authenticated outbound HTTPS/WSS. Only its credential hash is stored on the coordinator. Never enter a wallet private key or seed phrase. The control panel is loopback-only; shared native stages use private process pipes. There is no public inbound GPU port, remote shell or automatic background startup. Stopping closes the owned stage and removes its available shared capacity. OTHER SERVERS / MODELS The existing Node.js 20+ client can use an approved local llama.cpp or vLLM model. Keep the model endpoint on loopback. Match its exact model name to enrollment. Create a private worker.json and worker.token, then run: node ai/worker.mjs /private/path/worker.json Example config: { "coordinator": "https://limonata.xyz", "modelBase": "http://127.0.0.1:8000/v1/", "model": "exact-registered-model-name", "tokenFile": "/private/path/worker.token" } For an authenticated model endpoint, add modelKeyFile with a private local path. Only outbound HTTPS is required; never expose the model directly to the Internet. PAYMENT AND PRIVACY The quote records its selected model. The deposit binds the requester, provider payout, price, expiry and salted request commitment. The coordinator credits the provider after accepting the result, or refunds failed work. Expired requests can be refunded through the contract even if the coordinator is offline. Providers withdraw contract credits using withdraw(payoutAddress), paying any required transaction gas. The coordinator is trusted: a hash is not proof that a model ran correctly. There is no guaranteed utilization, wage or earnings rate. Shared Qwen3 32B is an explicit model choice and requires all stages to be ready. Independent whole-model routes remain separately identified. Every request shows its actual model; explicit provider/model choices stay pinned. Sponsored and paid access are separate choices; losing sponsored capacity does not silently switch a request to paid access. Community paid requests currently cost 0.01 test LIMO plus network transaction fees; the wallet confirms the actual quote. Testnet tokens have no monetary value, and this price does not set mainnet prices. The CPU demonstration remains a fallback. Users' messages are visible to the coordinator and selected provider. Do not log prompts or responses. Model weights/runtime are public downloads; protect your provider token and never commit it or wallet keys to source control. MAINNET Mainnet requires separate configuration, contract, chain ID, funded coordinator, validated providers and contract review. Testnet state/secrets must not be copied. The Windows preview cannot connect to mainnet without a separately reviewed release. COMPUTE DASHBOARD AND COLLECTION The app shows preparation/standby/computation state, accepted requests, output tokens, historical throughput and cumulative settled test-LIMO earnings. A collective throughput sample belongs to the whole group, not one GPU. Historical samples are not live speed; the current protocol does not measure live tokens/s. These are coordinator accounting records, not independent proof of inference or blockchain hashes. Waiting and local setup checks do not earn coins. Recorded settled earnings are not the current withdrawal balance. Escrow credits can include both earned provider credits and chat refunds. The panel separately reads escrow credits and known distributor claimable amounts on the expected chain at one block, cached for up to 15 seconds. An unavailable read is shown as unknown, never as a zero balance. The balance belongs to the payout wallet and is shared by all machines using that address; do not add those displays. Individual providers collect them with withdraw(recipient) using their payout wallet; this also covers automatic single-GPU work. Members of a configured paid multi-GPU group collect with release(payoutAddress) on its immutable pool contract. The provider app never needs your wallet private key. Wallet confirmation and network gas are required for collection. Test LIMO has no monetary value. HISTORY — AUTOMATIC COLLECTIVE MODE (0.4.0) Version 0.4.0 introduced Collective model mode. A local Vulkan inference check and current memory profile replace manual operator approval. The dashboard and public AI page show available capacity, model loading and conversation slots. The first target is Qwen3-32B Q4_K_M. Two chats share one instance; others queue. That release's automatic groups served sponsored requests only. The current single-GPU paid route above supersedes that limitation; larger automatic groups remain sponsored. For isolation, memory planning and limits see COLLECTIVE.txt. VERIFIED WEIGHTS CACHE — LINUX 0.5.6 / WINDOWS 0.4.1 Updated clients cache eligible public model weights after their first collective load, then verify and reuse them locally at subsequent loads. Maximum 32 GiB by default, keeping 2 GiB filesystem space free. Linux: --cache-gib; Windows: cache selector before Start. Chat data is not written to this weights cache. A client update and one reconnect are required. Retain the existing provider identity. Cache hits save network transfer; they do not accelerate token generation.