What they've actually shipped

EXHIBIT E · SCALE, NOT ADOPTION

Forget claims and roadmaps — HuggingFace's own API is fully public, no auth required, and it counts real downloads. Four model checkpoints, published under Pearl's own org, with 98,894 combined downloads. One honest caveat before the number does any work: this almost certainly isn't AI-community adoption — it's a miner-participation proxy. The official mining setup passes the model ID straight to the miner container (docker run … vllm_miner pearl-ai/Llama-3.1-8B-Instruct-pearl), so every miner who follows the docs triggers a download whether or not they care about the model itself. The lopsided like-to-download ratios (62,273 downloads, 6 likes on the top model) fit that read better than organic interest. Read it as "how many machines have set up to mine," not "how many developers adopted the model" — still a real number, just a different one than it looks like at a glance.

Llama-3.1-8B-Instruct-pearlbase: Meta Llama 3.1 8B62,273
Feb 2026 · 6 likes
Gemma-4-31B-it-pearlbase: Google Gemma 4 31B15,971
Apr 2026 · 5 likes
Qwen3-30B-A3B-Instruct-2507-pearlbase: Qwen3 30B-A3B (MoE)15,004
Jun 2026 · 0 likes
Llama-3.3-70B-Instruct-pearlbase: Meta Llama 3.3 70B5,646
Apr 2026 · 3 likes
Combined downloads, 4 models98,894
Org: pearl-ai
Members10
Followers46
PlanTeam
VerifiedNo
Fully open API — no token required for any of this.
Still actively maintained
Jun 16vision support added
Jun 18MoE checkpoint published
Jun 28benchmark results added
Real commits, one engineer, all within the last month checked — not a one-time launch dump.
The MoE tie-in
ModelQwen3-30B-A3B
TypeMixture-of-Experts
PublishedJun 11, 2026
This is the actual checkpoint behind the "MoE hard fork" already tracked on Case Status — a real model, not just a protocol change.

The 2-for-1, in writing

EXHIBIT F · THE MODEL

Strip the mystery away and "duplex" is just the team's word for the trick at the core of the whole protocol: compute that mines and infers at the same time — one GPU pass earning twice. It isn't marketing fluff, either: the idea is formalized in a peer-reviewed economics paper by one of the biggest names in the field — who happens to be on Pearl's payroll. (The name's own backstory lives in the Lore Corner.)

The paper that put it in writing

Rafael Pass also gave the full public talk on this paper — produced and published by Pearl Research Labs. Here's the paper, and the exact moment he defines "duplex" in his own words.

PAPERarXiv · June 2026
The Economics of Proof-of-Useful-Work.
Sole author: Rafael Pass, Cornell Tech — one of the most cited names in cryptography and blockchain-consensus theory. The paper builds a full competitive-equilibrium model of a proof-of-useful-work economy and names Pearl Network directly as the real-world system it's modeling. Buried in the byline: "This work was done while consulting for Pearl Research Labs." Pass isn't an outside academic analyzing Pearl from a distance — he's a paid consultant, and this is the first document that puts that relationship in writing.
QUOTE · 16:32–16:44from the talk
"They can do this thing referred to as a duplex operation… a joint, a two-for-one operation. It participates in both markets at the same time — it co-produces both mining and inference, but maybe with some overhead."
Rafael Pass, defining the term live, lightly cleaned up from the auto-generated transcript for readability.

Source: "The Economics of PoUW — Rafael Pass," produced and published by Pearl Research Labs on YouTube. ▶ Hear this exact moment at 16:10–16:44

The terminology tell: this paper puts "duplex work" and "Duplexia" into peer-reviewed print for the first time — the formal economic terms for a chain where token rewards subsidize AI inference. The word wasn't coined for the paper, though: it's the team's own internal name for the protocol, older than the paper itself (that story's in the Lore Corner). The read: Pass formalized the team's existing jargon, not the other way around.

The model, in one picture

M
Solo mining
Produces security only
S
D
Duplex
Produces both, with overhead α, γ
S + AI
L
Solo learning
Produces "useful output" only
AI

RECREATED FROM THE TALK'S "CONJOINT PRODUCTION MARKET" MODEL, IN OUR OWN LAYOUT — NOT A REPRODUCTION OF PASS'S ORIGINAL SLIDE.