pearl-2:native lowest published mining overhead, 3.9%, was measured on DeepSeek V3.2, a mixture-of-experts model, across 8 H200s.
Mixture-of-experts certificates have been rejected on Pearl’s mainnet since block 91,630 (Jul 23), so that mode doesn’t mine today.
For a dense model, the same Pearl post gives 5.08% (Llama 70B, 4 H200s), and @WeinsteinOmri put current integer mining at 10 to 15% on Sep 25.
For the FP8 upgrade, still without a fork height, his newest figure is 1 to 5% end to end (Oct 8).
Mixture-of-experts certificates have been rejected on Pearl’s mainnet since block 91,630 (Jul 23), so that mode doesn’t mine today.
For a dense model, the same Pearl post gives 5.08% (Llama 70B, 4 H200s), and @WeinsteinOmri put current integer mining at 10 to 15% on Sep 25.
For the FP8 upgrade, still without a fork height, his newest figure is 1 to 5% end to end (Oct 8).
1
2
11
751
The hash function under Pearl’s mining.
The team wrote a GPU kernel that runs it about 370× faster than the BLAZE3 implementation, nearly 400× on large inputs, so hashing adds little to the AI work that mines $PEARL.
The team wrote a GPU kernel that runs it about 370× faster than the BLAZE3 implementation, nearly 400× on large inputs, so hashing adds little to the AI work that mines $PEARL.
0
0
13
863
pearl-2:native inference platform lists the precision each model is served at:
DeepSeek V4 Pro 0813, both DeepSeek V4.1 Flash models and Qwen3.8-27B run at FP8.
GLM-5.3 and GLM-5.3-Flash run at FP4.
Mining on Pearl's mainnet uses 7-bit integer math today, and the planned FP8 upgrade (still a draft, with no fork height) would move it to FP8.
Mining on FP4 is listed as an open problem on Pearl's own Polymath page.
DeepSeek V4 Pro 0813, both DeepSeek V4.1 Flash models and Qwen3.8-27B run at FP8.
GLM-5.3 and GLM-5.3-Flash run at FP4.
Mining on Pearl's mainnet uses 7-bit integer math today, and the planned FP8 upgrade (still a draft, with no fork height) would move it to FP8.
Mining on FP4 is listed as an open problem on Pearl's own Polymath page.
2
4
20
1,308
Pearl $PEARL Supply check, week 28:
373.54M PEARL emitted, 17.79% of the 2.1B cap.
18% is about 2,050 blocks away, around Nov 13 or 14.
373.54M PEARL emitted, 17.79% of the 2.1B cap.
18% is about 2,050 blocks away, around Nov 13 or 14.
0
0
12
739
Pearl $PEARL mining, week 28:
Network hashrate 79.82 EH/s (on a 200-block estimate; earlier weeks used 60-block reading).
Kryptex mined 1,619 of the last 7 days' 3,375 blocks, 48.0%, down from 51.2% in week 27.
Network hashrate 79.82 EH/s (on a 200-block estimate; earlier weeks used 60-block reading).
Kryptex mined 1,619 of the last 7 days' 3,375 blocks, 48.0%, down from 51.2% in week 27.
0
0
20
1,179
New network hashrate high for Pearl pearl-2:native : 106.0 EH/s on Nov 6.
Until now, this account called highs on lordofpearls' 60-block reading. From this post on, highs use prlscan's 200-block estimate.
On that measure, the previous high was 105.5 EH/s on Oct 27.
The window ackers estimate hashrate from difficulty and the average time between recent blocks. L
At the same block, prlscan's data reads: • 200 blocks: 106.0 EH/s • 60 blocks: 150.1 EH/s
Blocks arrive at random intervals, so a short window turns a run of fast blocks into a spike, while a longer one smooths it out.
That is why trackers show different numbers at the same minute: lordofpearls' chart read 119.12 EH/s for this block.
Until now, this account called highs on lordofpearls' 60-block reading. From this post on, highs use prlscan's 200-block estimate.
On that measure, the previous high was 105.5 EH/s on Oct 27.
The window ackers estimate hashrate from difficulty and the average time between recent blocks. L
At the same block, prlscan's data reads: • 200 blocks: 106.0 EH/s • 60 blocks: 150.1 EH/s
Blocks arrive at random intervals, so a short window turns a run of fast blocks into a spike, while a longer one smooths it out.
That is why trackers show different numbers at the same minute: lordofpearls' chart read 119.12 EH/s for this block.
1
1
31
4,250
Rafael Pass, whose paper set out what a 51% attack would cost on a network like Pearl, has a new math result.
Matrix multiplication is the core math of AI models, and of pearl-2:native mining. @OpenAI recently proved it can be done in fewer steps, without showing how. Pass's paper gives the actual method, checked with proof software.
Matrix multiplication is the core math of AI models, and of pearl-2:native mining. @OpenAI recently proved it can be done in fewer steps, without showing how. Pass's paper gives the actual method, checked with proof software.
here are my thoughts about OpenAI's amazing Mat Mul result. comments very welcome.
1
0
17
1,024
Pearl pearl-2:native - October 2026: the monthly record.
Network hashrate roughly doubled, 42.10 to 87.42 EH/s on a 60-block reading from Sep 30 to Oct 31. The node shipped five releases.
Inside: who mined the blocks, which GPUs the draft PIP-3 keeps, and the inference platform's catalog changes.
Network hashrate roughly doubled, 42.10 to 87.42 EH/s on a 60-block reading from Sep 30 to Oct 31. The node shipped five releases.
Inside: who mined the blocks, which GPUs the draft PIP-3 keeps, and the inference platform's catalog changes.
0
0
20
1,143
Since this post: the FP8 upgrade's proof code is now on its branch.
On Nov 3 and 4 @prlnet merged three pull requests into fp8-scheme, the branch that carries the upgrade: a cap on how much work a verifier can be asked to do, a smaller verifier circuit (2^18 to 2^17 gates, per its author), and simpler tile selection.
It is the first change to that branch since Oct 5. Three steps remain before a fork: the branch merged into the main code, a pearl-2:native mainnet block height, and a release.
On Nov 3 and 4 @prlnet merged three pull requests into fp8-scheme, the branch that carries the upgrade: a cap on how much work a verifier can be asked to do, a smaller verifier circuit (2^18 to 2^17 gates, per its author), and simpler tile selection.
It is the first change to that branch since Oct 5. Three steps remain before a fork: the branch merged into the main code, a pearl-2:native mainnet block height, and a release.
“When FP8?” comes up in most pearl-2:native threads.
Where it stands on Oct 28:
The FP8 upgrade (certificate V4, a hard fork) sits in pull request #311, open since Sep 9, with no mainnet block height in the code.
On Sep 25 @prlnet CEO said it was on testnet and expected to launch “within weeks”. Work on its proof system is still in progress, with prover commits as recent as Oct 25. The signal to watch is a fork height for mainnet.
7
4
39
3,980
How do you confirm a miner actually ran the computation, without running it again yourself?
A zk-SNARK is a compact proof that a computation was done correctly, quick to check. Pearl uses it to validate every $PEARL block.
A zk-SNARK is a compact proof that a computation was done correctly, quick to check. Pearl uses it to validate every $PEARL block.
2
0
15
750
Anthropic released FlashProve on Nov 2, a prototype that proves a committed model produced a given answer.
On Llama 4 Maverick, proving took about 72 times the GPU time of plain inference, and about 94 times at the highest load tested.
@prlnet co-founder Ilan @komargodski on how it relates to Pearl: the two prove different things. An inference proof says which model answered. Pearl's proof of useful work says the computation was done.
(@AnthropicAI 's post does not mention pearl-2:native )
On Llama 4 Maverick, proving took about 72 times the GPU time of plain inference, and about 94 times at the highest load tested.
@prlnet co-founder Ilan @komargodski on how it relates to Pearl: the two prove different things. An inference proof says which model answered. Pearl's proof of useful work says the computation was done.
(@AnthropicAI 's post does not mention pearl-2:native )
Big news from @AnthropicAI on verifiable LLM inference, with an exciting connection to Pearl’s proof of useful work.
Two orthogonal questions:
• Did the model you paid for actually produce this response?
• Was computational work performed?
Anthropic reports that generating inference proofs currently takes ~70–95× the GPU time of ordinary inference. Pearl’s PoUW, on the contrary, adds negligible overhead. These costs buy different guarantees. Proving correct inference doesn’t automatically establish proof of work and the other way around. But the two are deeply complementary. Verifiable inference could certify which model produced an answer, while Pearl׳s PoUW only certifies that a computation has been done.
As inference proofs get more practical, that combination could make AI easier to trust and more economical to provide.
1
1
13
970
Pearl Supply check, week 27:
366.17M pearl-2:native emitted, 17.44% of the 2.1B cap (Nov 2).
The 18% mark is about 5,000 blocks away, around Nov 14.
366.17M pearl-2:native emitted, 17.44% of the 2.1B cap (Nov 2).
The 18% mark is about 5,000 blocks away, around Nov 14.
0
3
15
670
Pearl pearl-2:native mining, week 27:
Network hashrate 84.33 EH/s (60-block avg)
Kryptex mined 1,724 of the last 7 days' 3,365 blocks, 51.2%.
Network hashrate 84.33 EH/s (60-block avg)
Kryptex mined 1,724 of the last 7 days' 3,365 blocks, 51.2%.
1
1
24
3,338
How fast is PEARL's supply growing?
Over the next year it comes close to doubling, then grows more slowly each year after.
PEARL has no halvings. The block reward shrinks with every block: ~2,207 pearl-2:native today, ~1,515 a year from now.
Supply growth at the 194-second block target: Next year: ~82% Year 2: ~32% Year 3: ~18% Year 4: ~12%
Over the next year it comes close to doubling, then grows more slowly each year after.
PEARL has no halvings. The block reward shrinks with every block: ~2,207 pearl-2:native today, ~1,515 a year from now.
Supply growth at the 194-second block target: Next year: ~82% Year 2: ~32% Year 3: ~18% Year 4: ~12%
1
1
34
2,121
Which GPUs can mine Pearl after the FP8 upgrade?
The draft spec, PIP-3, says the verifier must reproduce each FP8 computation bit for bit, so hardware is whitelisted: “this proposal supports the Hopper and Blackwell architectures”.
That covers the H100, H200 and B200, plus RTX 50-series and RTX PRO 6000 cards. RTX 30 and 40-series cards and the A100 are currently not on the list.
The upgrade has no mainnet date yet, and until it activates, mining stays on the INT7 protocol those cards run today.
The draft spec, PIP-3, says the verifier must reproduce each FP8 computation bit for bit, so hardware is whitelisted: “this proposal supports the Hopper and Blackwell architectures”.
That covers the H100, H200 and B200, plus RTX 50-series and RTX PRO 6000 cards. RTX 30 and 40-series cards and the A100 are currently not on the list.
The upgrade has no mainnet date yet, and until it activates, mining stays on the INT7 protocol those cards run today.
12
3
40
4,136