Straight answers.

It pools consumer GPUs into a peer-to-peer network that serves LLM inference through an OpenAI-compatible API and makes each response independently falsifiable. Under the Exact Inference Profile, honest execution is bit-identical on every conforming GPU regardless of vendor, so every response can be backed by a signed receipt that anyone can check by re-executing it and comparing hashes.

No, and we publish that number ourselves. Verified consumer-GPU inference costs a multiple of the cheapest centralized per-token price floor. Datacenter bandwidth, batching, and subsidized pricing are structural advantages no peer-to-peer network abolishes. Mantissa sells provable inference, not cheap inference. If you do not need the receipt, a commodity API will beat us on price.

A signed, falsifiable claim attached to each job: the model hash, an input commitment, an output commitment, and intermediate checkpoint hashes over the raw 32-bit logits. Because execution is bit-exact, re-running the job on any conforming device either reproduces every hash and corroborates the claim exactly, or establishes a mismatch with a single differing byte. Receipts are designed for epoch Merkle rollups on Solana so, while the model, profile, and committed evidence remain available, they can be checked later.

Treat the base network as not private. The node serving your request can read the prompt it processes at compute time. No consumer-GPU network can cryptographically prevent that today, and we will not pretend otherwise. The network provides transport encryption, contractual non-retention, and routing control including operator allow-lists; a datacenter TEE tier is a possible future opt-in. Sensitive-data users should plan accordingly.

Cross-vendor bit-exactness is validated today on NVIDIA and AMD consumer GPUs across three backends (CUDA and Vulkan). Serving nodes run through a single Vulkan backend for both vendors, and every card must pass the conformance self-test on its own hardware before it can serve. The profile is certified per hardware generation as evidence accumulates, not assumed. The alpha targets 12–24 GB cards for 8B–32B-class models.

Not yet. The network is in a closed alpha and the public node client is not yet available. Onboarding is also deliberately quota-gated: supply is admitted in step with demand so that operator earnings stay real rather than diluted. Join the waitlist from the Network page and you will get one email when onboarding opens.

No. $XIP is not live on any chain. Nothing has been minted, sold, or distributed. Issuance and transferability remain gated on a publicly demonstrable network plus legal, economic, contract, and operational review. Any token trading under the Mantissa or XIP name today is unaffiliated with this project. When $XIP launches, the contract address will be announced only on mantissa.network and the official channels linked in the footer. Verify there and nowhere else.

Two structural differences. First, verification: subjective or vote-based scoring cannot prove any individual result correct, and tolerance-band checking can be gamed. Mantissa's exact execution makes every result provable by hash comparison. Second, emissions: many DePIN designs reward hardware presence, which can attract supply without demand. $XIP has no emissions faucet. Minting happens only against receipts that verified demand paid for.

Open-weight models from a whitelisted registry, served whole per node, roughly 8B–32B-class quantized models at launch. The capability ladder then climbs: metro-area clusters for ~70B-class models, and wide-area sharding of frontier-scale mixture-of-experts models as a gated research track that enters engineering only if the transport analysis survives.

It can try. The production economics are designed and must be calibrated to make that attempt negative expected value. Receipts are spot-checked by sampled re-execution, auditors are drawn by verifiable randomness so a node cannot know which jobs will be checked, and under exactness a single differing byte proves a mismatch. The launch gate requires expected loss from detection to exceed the gain from serving cheaper work; referee-confirmed dishonesty enters the full-stake penalty path. In our validation, every model-substitution attempt was caught, in every trial.

A floating-point number is a sign, an exponent, and a mantissa. When two honest GPUs disagree, the disagreement lives in the mantissa bits. That disagreement is the reason decentralized inference could never be verified. Mantissa's kernel standard makes every mantissa bit identical on every vendor's silicon. The network is named for the bits it defends; the $XIP ticker is named for the Exact Inference Profile that defends them.

Start with the whitepaper and the technology page. Full kernel-level method details are documented in a U.S. provisional filed July 10, 2026 and reserved for a subsequent technical paper. Evidence records, including per-position output hashes, sweep results, and the economic model, are retained and will accompany that publication.