Pearl (PEARL) mining is built on matrix multiplication, the core operation of AI. But modern AI models are not one big matrix multiplication. Many of the largest ones use a design called mixture of experts (MoE), where the work is split into many smaller multiplications. PIP-2 is the Pearl Improvement Proposal that makes that kind of work cheap to prove. This guide explains what MoE is, why the original mining certificate handled it poorly, and what the version 2 certificate changes.
A quick recap of Pearl mining
Pearl uses proof of useful work. A miner runs noised matrix multiplications on a GPU. The noise is tied to the block being mined, so the result doubles as a lottery ticket. When a ticket hits the difficulty target, the miner attaches a zero-knowledge proof (built on Plonky2 and STARKy) and other nodes check that proof instead of redoing the math. The construction comes from a 2025 paper by Ilan Komargodski and Omri Weinstein, arXiv 2504.09971.
The package of commitments and proof that a miner submits is called a certificate, and it has a version number. Version 1 proves one matrix multiplication. PIP-2 adds version 2.
What is a mixture-of-experts model?
In a normal (dense) layer, every token of input passes through the same weight matrix. In a mixture-of-experts layer there are several weight matrices, called experts. A small router looks at each token and sends it to only a few of them, the top k out of n experts. The model gets the capacity of many experts while each token only pays for a few.
For the hardware, that turns one large multiplication into many smaller ones, one per expert. PIP-2 gives the average number of tokens each expert sees as:
tokens per expert = M × top_k / n_E
where M is the number of tokens, top_k is how many experts each token visits and n_E is the number of experts. Running all those expert multiplications together is known as a grouped GEMM (GEMM is the standard name for a general matrix multiply), which is where the proposal gets its name.
Why version 1 was a poor fit
Version 1 certificates were designed around one matrix multiplication at a time. To mine on MoE work with version 1, a miner would have to treat each expert as a separate job. According to PIP-2, that means repeating the commitment hashing, the noising and the final cleanup step (the spec calls it peeling) for every expert, over the same activations again and again. The extra overhead eats into the “useful” part of useful work, which is supposed to cost very little on top of the real computation.
What the version 2 certificate does
PIP-2 restructures the proof around the whole MoE layer instead of each expert:
- Commit once to the activations. The miner commits a single time to the global activation matrix A, the inputs for every token.
- Commit once to the experts. The miner commits a single time to the stacked expert-weight matrix W, all the experts laid on top of each other.
- Commit once to the routing. The miner commits to the routing table, the record of which token went to which experts.
- Noise once, reuse everywhere. The activations are noised a single time before routing, and that noised copy is reused across all the experts.
On the checking side, the verifier confirms the opened tile of the result as before, and also checks that the opened rows match the routing table and the committed activations. That stops a miner from claiming a routing that did not happen.
Version 1 vs version 2 at a glance
| Version 1 | Version 2 (PIP-2) | |
|---|---|---|
| Proves | One matrix multiplication | A grouped set of expert multiplications |
| Commitments on MoE work | Repeated per expert | Once each for activations, experts and routing |
| Noising on MoE work | Repeated per expert | Once, before routing |
| Dense matrix multiplication | Yes | Yes, as a special case with virtually the same overhead |
| Arithmetic | Integer (INT7 × INT7 to INT32) | Integer (INT7 × INT7 to INT32) |
Details miners should know
- A version 1 win is not a version 2 win. PIP-2 notes that a winning version 1 witness is generally not a winning version 2 witness. The two are separate lotteries over the same kind of math, so software has to target one or the other.
- Version 1 keeps working for non-MoE miners. The proposal says miners not running MoE workloads may continue with version 1, subject to the activation rules.
- The router does not count. The small matrix multiplication the router performs is not counted as useful work. PIP-2 treats it as negligible.
- It is a hard fork. PIP-2 is a Standards Track, Consensus proposal created on June 1, 2026, and its status in the PIPs repository is Final. Nodes need software that understands version 2 certificates. Check the official releases page for the node version you should run; as of November 5, 2026 the latest release is pearld v1.5.5.
How PIP-2 relates to the FP8 fork (PIP-3)
Both version 1 and version 2 certificates prove integer matrix multiplications. PIP-3, a proposal still in Draft as of November 5, 2026, would move Pearl mining to floating point (FP8 × FP8 into an FP32 accumulator) with a new version 3 certificate. As written, PIP-3 says that at activation version 3 becomes mandatory and versions 1 and 2 are rejected. No activation height has been set.
So PIP-2 is best read as one step in a direction: making the proof fit the way real AI models are built. PIP-2 adapted the certificate to MoE layers; PIP-3 would adapt the number format. Our explainer on the FP8 fork (PIP-3) covers that change, and how Pearl Improvement Proposals work explains the Draft, Proposed and Final statuses.
Does this make Pearl mining more useful?
It removes a cost that would have discouraged mining on real MoE inference. It does not, by itself, prove that the matrices came from a real model: the protocol checks that the multiplication is correct, not where the inputs came from. Whether the network’s hashpower does real AI work is an open debate, which we cover in is Pearl mining useful?
Where to go next
- New to mining? Start with how to mine Pearl and Pearl mining pools compared.
- Choosing hardware? See the best GPUs for Pearl mining.
- Need a payout address? Oyster’s miner mode gives you one fixed prl1p address for your pool. See the mining payout wallet guide and the mining docs.
Frequently asked questions
What is PIP-2 in Pearl?
PIP-2, titled Grouped-GEMM Proof-of-Useful-Work for Mixture-of-Experts, is a Pearl Improvement Proposal that adds a version 2 mining certificate. It lets a miner prove the expert matrix multiplications inside a mixture-of-experts AI model efficiently, instead of proving each expert separately. Its status in the PIPs repository is Final, and it is a hard fork.
Do I have to switch to version 2 certificates to keep mining?
Not under PIP-2 itself. According to the proposal, miners that are not running mixture-of-experts workloads may continue to use version 1 certificates, subject to the activation rules. Note that PIP-3, the proposed FP8 fork, would reject both version 1 and version 2 at its activation as currently written.
Can a version 2 certificate prove an ordinary dense matrix multiplication?
Yes. PIP-2 says a dense matrix multiplication is a special case that can be proven with version 2 at virtually the same overhead as version 1.
Does PIP-2 change how much PEARL a block pays?
No. PIP-2 changes the format of the mining proof, not the emission formula. The block reward still comes from the same per-block formula described in our supply guide.
