Token-less decentralized personal-computer compute network based on useful-work credits, demand pooling, and compute-offset accounting.
中文暂定名：协同算力网
CoCompute Network does not issue tokens and does not build a speculation layer.
It focuses on one thing: turning idle personal-computer compute into non-transferable usage credits that reduce future AI and compute costs, while also grouping overlapping demand when multiple users need the same or similar result.
-
Useful-work contribution
Users can opt into a lightweight worker that contributes compute only under local resource limits. The network favors work that is verifiable, useful, and privacy-aware.
-
Demand pooling
When multiple users submit identical or highly similar tasks within a short time window, the gateway groups them into one core workload.
-
Peer-to-peer computing
Participating PCs contribute local GPU or CPU resources and form a temporary heterogeneous compute cluster to complete the workload together.
-
Compute-offset accounting
In a normal API workflow, every user pays the full cost independently. In CoCompute Network, users who contribute more compute receive a larger offset against their own model or token usage cost.
For example, if a participant provides 60% of the effective compute for a pooled task, most or all of their cost for that task can be offset. The remaining cost is shared by nodes that contributed less compute.
For short chat-style requests, real-time contribution is usually too small to matter. The practical model is long-lived contribution: users earn non-transferable credits while their device is idle, then spend those credits later on AI usage.
Most decentralized compute networks are markets: providers sell compute, requesters buy compute, and the network often relies on a tradable token.
CoCompute Network is closer to a cooperative cost-sharing system:
- No native token.
- No speculative incentive.
- High-value credits are preferably bound to work the contributor needs or will later use.
- Credits exist to offset usage, not to become a financial asset.
- The system prioritizes answer-bound and cohort-bound work over open-ended mining.
- Token-less by default.
- Useful work only.
- Open-source first.
- Local-first participation.
- Verifiable contribution accounting.
- Privacy-aware prompt grouping.
- No hidden monetization through compute mining.
Groups similar prompts or compute requests using lightweight embeddings, locality-sensitive hashing, and task fingerprints.
Open questions:
- How similar is similar enough?
- How do we avoid merging prompts with subtly different intent?
- Can users opt into strict, fuzzy, or private pooling modes?
Coordinates heterogeneous personal devices for a single pooled task.
Possible directions:
- Layer or tensor sharding for LLM inference.
- Batch-level or shard-level execution for backtests and data jobs.
- Fallback execution when a peer drops.
- Local sandboxing for untrusted workloads.
Measures useful compute rather than claimed hardware capability.
Possible signals:
- Verified completed work units.
- Runtime-normalized throughput.
- Device benchmark calibration.
- Redundant spot checks.
- Peer attestation or lightweight proof-of-execution.
Not all contribution should be treated equally. CoCompute uses three credit classes:
-
Self-bound work
The contributor's device performs work that directly affects their own current or future answer. This receives the highest offset rate because poisoning the work harms the contributor.
-
Cohort-bound work
A group of users with overlapping demand contributes to a shared result. This receives a high or medium offset rate because participants still care about the result.
-
General verified work
Idle devices perform generic tasks for the network. This receives a lower offset rate and requires stronger verification because the contributor may not care about the final answer.
Records task participation and usage offsets without creating a tradable token.
Possible shape:
- Git-like hash chain.
- Local-first replicated event log.
- Signed receipts for each pooled task.
- Non-transferable account credits.
The strongest incentive alignment comes from binding high-value work to answers the contributor needs.
For example, a user who plans to analyze a repository tomorrow can let their device precompute file indexes, embeddings, dependency graphs, and static-analysis artifacts tonight. If they poison that work, their own answer gets worse.
General offline contribution is more vulnerable to poisoning because the worker may only care about earning credits. For this reason, generic work should be limited to tasks that can be verified cheaply, such as deterministic batch jobs, embedding work, redundant validation, fixed-seed model calls, or benchmark-style work units.
However, the system still needs engineering defenses against:
- Sybil accounts.
- Low-effort freeloading.
- Malicious partial results.
- Prompt poisoning.
- Privacy leakage through semantic grouping.
- Fake contribution reports.
The first prototype should assume partial trust and use redundant verification before attempting fully permissionless operation.
See docs/incentives-and-poisoning.md for the current incentive model.
See ROADMAP.md.
This project is inspired by several adjacent systems, but aims at a different combination:
- Volunteer computing systems such as BOINC.
- Distributed AI systems such as Petals and Hivemind.
- Community AI compute systems such as AI Horde.
- Decentralized compute markets such as Golem, Akash, and Render Network.
- LLM serving techniques such as batching, prefix caching, semantic caching, and KV-cache reuse.
CoCompute Network's intended niche is token-less cooperative cost reduction for users with overlapping demand.
Concept and early architecture draft. Contributions, critiques, and prototype experiments are welcome.
MIT. See LICENSE.