a16z "Disciple" Inference Practical Guide Three: Epoch3 Incentive Testing Launched, Multicoin Capital Invests Heavily

Written by: J1N, Techub News

Epoch3 officially launched on June 7, 2025, introducing significant protocol changes including technical improvements, a new staking mechanism, and operational transformations.

Kuzco has undergone a brand upgrade and is now renamed Inference.

The news that the project has received investment from Multicoin Capital is true.

The entry threshold for devices has increased.

Launch a staking mining mechanism similar to io.net.

Epoch2 Review

Recommended configuration for participation

It is recommended to review previous articles before reading.

"a16z 'Disciple' Kuzco Practical Guide: How to Efficiently Mine AI Computing Power?"

"a16z 'Disciple' Kuzco Practical Guide II: From Solo Operations to Cluster Deployment"

Multicoin Capital Enters the Game

In December of last year, Inference founder Sam claimed in the official DC Gold Miner channel that the project had received $11.5 million in funding from Multicoin Capital and a16z csx. This news has recently been confirmed, as the project is listed in the Portfolio of Multicoin Capital. It is true that Inference received investment from Multicoin Capital. (Multicoin Capital is also an early investor in io.net.)

Epoch 3 Early

Since the launch of Epoch 3 on June 7, 2025, the network has been running for 10 days. Compared to the previous two phases, the number of participating miners has significantly increased, and both the inference volume and network stability have improved significantly. As of the time of writing, the number of workers has reached 12,100. Although this number has not yet surpassed the peak of 14,000 in Epoch 1, it is reasonable to infer that the current number of participants is several times that of the early stages, as Epoch 3 has restricted low-power graphics cards and multiple instances.

What changes does Epoch 3 bring?

Automatic Node Update

The automatic node update feature can significantly reduce the operational burden on miners. After experiencing Epoch 1&2, it was found that the official frequently updates files irregularly, and the update notifications are not timely. This has led many users to terminate processes, thinking that there is a problem with the device, wasting a lot of time troubleshooting, while the actual reason is that users did not update in a timely manner.

Unified Reasoning Engine Management

In both rounds, the author ran Meta's Llama-3.1-8B model and did not see the options for which models can be selected as stated in the official documents. This is questionable. Additionally, if options are available, it is preferable to choose models with higher usage rates, as the inference volume is larger and the scores obtained will be higher.

Enhanced GPU Detection and Verification

Inference announced as early as Epoch 2 that it would forcibly remove graphics cards with computing power below RTX 3080. It is estimated that they were concerned about dissatisfaction in the community, which is why it has not been implemented until the later stage of Epoch 2. It can still be seen that there are a large number of insufficient computing power graphics cards like 3060 and 3070 participating in inference. The author believes that devices with insufficient computing power will severely affect the user's experience, and eliminating these graphics cards can bring a qualitative improvement to the entire product.

Inference Chinese DC Community

Equity-weighted job routing

Equity-weighted job routing prioritizes the points earned by miners and adds a k parameter to the formula for calculating the score per unit of work.

Priority Score = 1 + k * (Miner Device VRAM / Total Network VRAM * Total Network INT Staking Amount * Miner Reputation Weight)

When k = 0: The routing adopts a circular method, providing equal value points for all miners.

When network utilization is low: the k value will be increased, which will increase miners' rewards.

When network utilization is high: the k value will be reduced, causing miners of various scales to have their rewards tend to be balanced.

By dynamically changing the k parameter, the most reasonable incentives and optimal resource utilization can be achieved under different working conditions of the network. For miners, this means that even during periods of low demand, there are still decent point rewards. During peak demand periods, even miners with minimal stakes can contribute and receive rewards.

Dual Token System

Epoch3 has launched a dual-token system, INT points and INT-DEV tokens. Currently, the test tokens and points have no value and are only used for testing.

INT points are mainly used to calculate the workload of miners, and they are also an important indicator of network participation at the current stage.

INT-DEV token is the token for the Solana Devnet test network and has no value. It is mainly used for airdrop and reward distribution testing, and its current role is to test the staking system.

Staking System

The staking system uses the SPL token standard of the INT-DEV token. This system is similar to an accelerator, allowing any miner to create an INT staking pool, set the commission rate, and attract other INT holders to stake.

As the creator of the staking pool, the more people stake and the more INT they have, the more inference tasks they can allocate from the network. The pool owner sets a commission rate when creating the pool. After each inference task is completed, points will be awarded to the staking pool, and after the pool owner withdraws their share, the remaining points will be distributed to the staking users in the pool.

As a user with only INT tokens and no mining machines, you can stake your INT in high-yield mining pools to obtain better returns. The goal is to find pools with high machine computing power and fewer stakers.

The author's staking pool welcomes everyone to stake. Currently, there are no earnings from staking; it is for testing purposes only.

Here we can see the shadow of Multicoin Capital's guidance, which further confirms its participation in the investment. The staking mechanism of Inference is similar to that of io.net in expanding the investor base. This is considered an advantage for slow-progress projects, as the model developed by leading projects in the same sector can be improved and utilized. However, this does not mean that the introduction of the staking mechanism will have a positive impact on token prices; the performance of IO is a clear example.

Reputation System (to be launched in late Epoch 3)

The reputation system assigns credit values based on the performance of miners, evaluating their inference throughput and operational stability. The author believes that this mechanism can promote the decentralization of the project, which is much stronger than many projects that purely sell nodes or allow participation in inference simply by paying. It truly is a project that focuses on real work.

Epoch 2 Review

Epoch2 started in November last year, and the early performance of Epoch2 was quite poor. In the first three months, the total network inference volume was relatively low, only 10-20% of the usual amount. Regarding this, the official response before the second phase started in DC stated, "The simple answer right now is that the points will be converted in a reasonable way, considering their value relative to other parts of the network at the time of acquisition." This ensures that early participants receive appropriate rewards while also considering that we need to continue to incentivize operators to contribute their computing power. In other words, incentives will be distributed relatively fairly to participants based on the actual network operation conditions.

The author makes a reasonable inference about the poor performance in the second phase:

At that time, the official promise was to remove graphics cards below RTX3090, but in reality, the official did not do so, resulting in many tasks being assigned to GPUs with insufficient computing power, such as RTX3060. This leads to some issues, such as when the task volume is limited, the task is obtained by RTX3060, but RTX3060 has a slow inference speed, ultimately causing high-power cards like RTX3090 and RTX4090 to not receive tasks, thus naturally having no inference volume. This leads to a decrease in scores.

On the other hand, the Inference team (formerly Kuzco) participated as an important partner in Solana's AI hackathon last December, coinciding with the time when the network had issues. It is reasonable to speculate that Sam and the Inference team focused their efforts on the hackathon and did not maintain the platform well.

This situation continued until mid-February, when the network began to return to normal. However, after returning to normal, the number of graphics cards participating in mining was far less than in Epoch 1. Additionally, the amount of tasks received by each card was also much less compared to Epoch 1, because the official limited multi-opening in Epoch 2. According to the official website's rules, one GPU can only run one worker. In practice, it is possible to run multiple instances on a single card. The author has previously open-sourced the multi-opening script on GitHub.

Configuration Recommendation

The following is a configuration combination with a relatively high cost-performance ratio based on the author's testing: X99 + E5 + RTX3090. Previously, due to the tariff war, the price of the 3090 graphics card rose from an average of 5700 yuan to 6700 yuan. Now the situation has improved, the market has cooled down, and the price of the graphics card has dropped to pre-tariff levels, making it a good time to purchase. Additionally, if readers want to invest in component mining machines for this project, they will initially incur a loss of 20-30% due to equipment depreciation costs, with electricity charges calculated separately. If you want to choose a cloud computing power supplier or intermediary service provider, you should ensure that they have the capability to cope with the instability of startup projects.

Lastly, a reminder: Inference is an early AI mining project and currently has not disclosed its funding for unknown reasons. The network's operation is not stable enough, often experiencing network downtime, and frequent updates without notice lead to miners going offline, among other issues. Another risk is the unknown returns; currently, only points can be earned. Whether it is worth investing manpower and resources to participate in this project, please consider it carefully.

Motherboard: X99 Dual U Multi-card Direct Insert Platform

CPU: E5 2680V

Memory: 32GB+ (Multiple cards starting simultaneously will temporarily occupy a large amount of memory)

The power supply is determined by the power consumption, for example, a 6-card 3090 configuration requires dual power supplies (you need to configure a power supply parallel startup line).

Hard disk: 500GB+ (a separate AI model needs to be downloaded for each process, which has certain requirements for the hard disk)

Network: Gigabit or above (the network has a significant impact on work and needs to be well configured)

6 cards 3090 platform, single machine full load 3-4kw, actual power consumption 1-2kw. (Mainly depends on network operation conditions, and not running at full power 7*24 hours.)

Budget: Mainboard ¥700, CPU ¥200, Power Supply ¥600, 6*3090 ¥36,000, Chassis and other configurations ¥600. A complete set costs about ¥38,000. The computing power when running the meta llama 8b model is approximately 600 Toks/s, with a theoretical daily inference volume of 50M per unit, and actual measurements around 10-20M. This data is for reference only.