The BlockZero Model Marketplace: What Phase 2 Looks Like
Phase 1 of BlockZero is a training platform: customers bring domain data, miners train expert modules, validators evaluate quality, the expert library grows. Phase 2 is what the library enables: a marketplace where validated domain experts are discoverable, testable, and licensable.
What the Expert Library Becomes
Every training job in Phase 1 deposits a validated expert module into the library. By the time we reach Phase 2, the library contains experts across dozens of domains — legal reasoning, financial analysis, medical literature, scientific domains, code generation specializations, language-specific experts, and more.
Each expert module comes with:
- Benchmark performance data: measured on standard and custom benchmarks relevant to the domain
- Training provenance: what data was used, how many training cycles, which base model, which miners contributed
- Live demo: run inference on the expert directly in the browser before committing to a license
The marketplace is the storefront for this library. It's where enterprise buyers browse, compare, and test experts before deploying them.
Figure: Conceptual layout of the BlockZero Model Marketplace — browse by domain, compare benchmark performance, and test with custom prompts before licensing.
How Expert Licensing Works
An enterprise buyer who needs a financial analysis expert has two options in Phase 1: pay for a training job using their proprietary data, or wait for someone else to train a general-purpose financial expert and use that. Phase 2 adds a third option: license a validated expert that was already trained.
Usage-based licensing: pay per token routed through the expert. The expert is served by BlockZero's inference infrastructure; the buyer doesn't need to manage model weights or deployment.
Weight licensing: for buyers who want to deploy the expert in their own infrastructure. They receive the expert weight files, the router calibration, and integration documentation. One-time payment plus a per-year maintenance fee.
Exclusive domain licensing: for buyers who want to ensure their domain expert is not accessible to competitors. They pay a premium to make the expert private (removing it from the marketplace) and optionally fund continued training cycles to keep it current.
Revenue from expert licenses flows back to the miners who trained those experts. This closes the economic loop: miners who train high-value domain experts earn not just from the initial training reward, but from ongoing licensing revenue for as long as the expert is deployed.
The TAAS Factory Model
We think about Phase 2 as "Training-as-a-Service Factory" — a system that doesn't just sell access to trained models, but produces purpose-built models on demand.
The factory metaphor is deliberate. A factory has:
- Inventory (the expert library): validated components ready to ship
- Production lines (the training subnet): capacity to produce new components on demand
- Quality control (Proof-of-Loss + validator evaluation): every component verified before it ships
- Assembly (WiSE-FT, Router Annealing): components integrated into the base model correctly
A buyer who comes to the marketplace with a highly specialized need — say, a model for analyzing FDA regulatory submissions — can either:
- Find an existing expert that partially covers their domain and license it immediately
- Commission a training job using their proprietary data, producing a custom expert that's added to their private library
Option 1 is fast and cheap (the library has done the work already). Option 2 is slower and more expensive but produces a model precisely calibrated to their specific data. In many cases, Option 1 provides a near-specialist starting point that Option 2 can then refine — compounding the investment.
How This Compares to HuggingFace Hub
The HuggingFace Hub is the obvious comparison: a repository of pre-trained models, community-contributed, with benchmarks and demos.
| HuggingFace Hub | BlockZero Marketplace | |
|---|---|---|
| Quality guarantee | None (community uploaded) | Proof-of-Loss validated |
| Custom training | Not provided | On-demand via subnet |
| Provenance | Self-reported | On-chain attestation |
| Revenue for contributors | None | Licensing revenue share |
| Composability | Manual merging | Native MoE integration |
| Specialization depth | Dense fine-tuned models | Expert module + base model |
The key difference is the quality guarantee and composability. HuggingFace hosts fine-tuned dense models, which are point-in-time artifacts with no ongoing quality guarantee. BlockZero marketplace experts are:
- Validated by Proof-of-Loss (quality at training time is measured, not claimed)
- Composable with the shared base model (multiple experts can be combined without conflict)
- Live assets that can be retrained as domain data evolves
Miner Incentives in Phase 2
Phase 2 changes the miner reward structure substantially.
In Phase 1, miners earn from training cycle rewards only — a per-cycle TAO payment proportional to their Proof-of-Loss score. The reward stops when the cycle ends.
In Phase 2, miners earn from both training cycle rewards and ongoing licensing revenue from experts they contributed to. An expert that becomes heavily used in production generates licensing fees for years after the training cycle that created it.
This creates fundamentally different time horizons for miner strategy:
- Short-term thinking (Phase 1): maximize per-cycle Proof-of-Loss score
- Long-term thinking (Phase 2): prioritize training experts in domains with high expected licensing demand
Miners who correctly anticipate high-demand domains and train experts for them early — before the domain becomes contested — earn outsized long-term revenue. First-mover advantage in domain coverage is economically significant.
This also creates alignment between miners and expert library quality: miners have strong incentives to produce experts that remain valuable over time, not just experts that pass the immediate Proof-of-Loss evaluation. A mediocre expert that passes evaluation but performs poorly in production earns no licensing revenue. A high-quality expert earns consistently.
The Flywheel, Completed
Phase 1 creates a flywheel: more customers → more training jobs → richer expert library → better starting baselines → lower training costs → more customers.
Phase 2 extends the flywheel: richer expert library → marketplace with discoverable, licensed experts → faster time-to-deployment for enterprise buyers → more use cases served → more revenue → more training investment → richer expert library.
The two flywheels reinforce each other. The marketplace creates additional demand for library depth (buyers want more domain coverage). Additional demand drives more training jobs. More training jobs improve library quality. Better library quality attracts more marketplace buyers.
The moat deepens with every cycle. An expert library with 18 months of accumulated domain coverage across dozens of fields is not something a new entrant can replicate quickly — it represents thousands of training cycles, millions of dollars of compute, and iterative quality improvement from many rounds of Proof-of-Loss evaluation.
Timeline
The marketplace is a Phase 2 commitment — we will not ship it until the Phase 1 library is deep enough that the marketplace has something worth browsing.
Our internal threshold: the library should contain validated experts across at least 20 distinct domains, with at least 3 training cycles of iteration per domain, before the marketplace opens. We expect to meet this threshold in Q3-Q4 2025 depending on network adoption.
When it's ready, we'll announce it. Until then, we're focused on building the library that makes the marketplace worth visiting.