Independent deployments will determine whether Lumabri and Colibri can serve mixture-of-experts models across peer-to-peer commodity machines with practically useful throughput and reliability.

state: expiredheat: lowuncertainty: highknownscott: mediumdistributed-inference moe-models local-inferenceLumabriColibriJustVugg

What is this?

Colibrì is an open-source inference engine from JustVugg that runs very large mixture-of-experts models on commodity hardware by staging only the routed experts across VRAM, RAM, and storage instead of keeping the full model in fast memory. Lumabri is presented as a peer-to-peer layer for distributing weights and executing experts across multiple machines, but the supplied snippets provide little independent detail about its implementation or measured reliability. Available reports characterize Colibrì as a proof of concept with a steep performance penalty—potentially useful for offline workloads, but currently too slow for ordinary real-time chat—and do not establish that independent Lumabri deployments have achieved practical throughput or reliability.

Why it matters to Scott

The radar already tracks this exact unresolved development in `radar:lumabri-peer-to-peer-moe-inference`, including the need for independent throughput and reliability testing. It still bears directly on Scott’s hardware-aware local-inference work and self-hosted GPU substrate—and echoes his ChessBrain distributed-computing experience—but the supplied evidence does not yet add a validated result that would change his builds or position.

dev:concept.hardware-aware-local-inferencedev:project.gamepcwork:project.chessbrain-netip:concept.ai-unit-economicsradar:lumabri-peer-to-peer-moe-inferenceradar:concept.distributed-inferenceradar:concept.moe-inferenceradar:concept.local-inference

queries asked of Scott's wikis

peer-to-peer distributed inference across commodity machines
mixture-of-experts expert routing and weight streaming
local inference memory hierarchy and storage economics
distributed inference reliability and heterogeneous peers
model sovereignty through pooled consumer hardware
benchmarks for practical local-model throughput

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (2) — ⭐ canonical anchor

source	object	author	score	comments
🟧 hn	Show HN: Lumabri – Run Moe Models on a P2P Swarm with Colibri	vforno	44	19
🟧 echo.github ⭐	The first substantive Lumabri commit describes “P2P weight distribution” and peers executing experts, with identical-token measurements. The	Vincenzo Fornaro	—	—

Interpretation history

2026-08-19T13:29:59Z
The announcement has produced no independent deployment, benchmark, or reliability evidence across repeated checks, and its discussion has stopped developing. Retire active monitoring until a concrete heterogeneous-peer implementation or measured result creates a new episode.
2026-08-17T12:43:47Z
The staleness check found no new deployment, benchmark, or reliability evidence. The case remains an unvalidated implementation announcement, with any practical peer-to-peer MoE utility still speculative.
2026-08-15T12:28:34Z
The refreshed discussion adds only hypothetical LAN and household-node applications plus concern about inference latency; it still supplies no independent deployment, benchmark, or reliability result. This is repetitive amplification of the implementation announcement, not validation.
2026-08-14T13:41:19Z
The refreshed comments only speculate about LAN and low-cost-device use cases while leaving verification, throughput, and reliability unanswered. No independent deployment or benchmark changes the implementation-announcement status.
2026-08-14T10:35:30Z
The refreshed discussion remains speculative and adds no independent deployment, benchmark, verification method, or heterogeneous-peer reliability evidence; the case is still an early implementation awaiting practical validation.
2026-08-14T08:40:07Z
The reobservation adds only minor engagement and no independent deployment, benchmark, or reliability evidence. The case remains an implementation announcement awaiting practical validation.
2026-08-14T08:31:03Z
grounded: known/medium — The radar already tracks this exact unresolved development in `radar:lumabri-peer-to-peer-moe-inference`, including the need for independent throughput and reli
2026-08-14T08:28:51Z
origin walked (codex/luna, conf 0.98): anchor hn.story.49293523 -> echo.github.0ad192bd9b by Vincenzo Fornaro
2026-08-14T08:27:41Z
case created — The first-party repository is a concrete implementation of a distinct peer-to-peer architecture for distributing sparse-model inference.

Decision trace

08-19 23:30expireThe announcement has produced no independent deployment, benchmark, or reliability evidence across repeated checks, and its discussion has stopped developing. Retire active monitoring until a concrete
08-19 23:29alert_silentThe only delta is a small score increase without new comments or technical evidence; alerting would add no decision-relevant information for Scott.
08-19 23:29alert_routeThe only delta is a small score increase without new comments or technical evidence; alerting would add no decision-relevant information for Scott.
08-17 22:43repriceThe staleness check found no new deployment, benchmark, or reliability evidence. The case remains an unvalidated implementation announcement, with any practical peer-to-peer MoE utility still speculat
08-17 22:43alert_silentNo consequential delta occurred; this can wait until an independent deployment or measured heterogeneous-peer result appears.
08-17 22:43alert_routeNo consequential delta occurred; this can wait until an independent deployment or measured heterogeneous-peer result appears.
08-15 22:28repriceThe refreshed discussion adds only hypothetical LAN and household-node applications plus concern about inference latency; it still supplies no independent deployment, benchmark, or reliability result.
08-15 22:28alert_silentNo consequential new fact has emerged; the added comments remain speculative, so waiting for an independent deployment or measured heterogeneous-peer result carries little attention regret.
08-15 22:28alert_routeNo consequential new fact has emerged; the added comments remain speculative, so waiting for an independent deployment or measured heterogeneous-peer result carries little attention regret.
08-15 00:21sensor_dirtyengagement_update
08-14 23:41repriceThe refreshed comments only speculate about LAN and low-cost-device use cases while leaving verification, throughput, and reliability unanswered. No independent deployment or benchmark changes the imp
08-14 23:41alert_silentThe discussion adds questions and hypothetical applications rather than consequential evidence; this can wait for the next briefing pending an independent deployment or measured heterogeneous-peer res
08-14 23:41alert_routeThe discussion adds questions and hypothetical applications rather than consequential evidence; this can wait for the next briefing pending an independent deployment or measured heterogeneous-peer res
08-14 22:21sensor_dirtycomment_update
08-14 21:21sensor_dirtycomment_update
08-14 20:35repriceThe refreshed discussion remains speculative and adds no independent deployment, benchmark, verification method, or heterogeneous-peer reliability evidence; the case is still an early implementation a
08-14 20:35alert_silentThe new comments are praise, jokes, and an unanswered verification question rather than consequential evidence, so waiting for the next briefing carries no attention regret.
08-14 20:35alert_routeThe new comments are praise, jokes, and an unanswered verification question rather than consequential evidence, so waiting for the next briefing carries no attention regret.
08-14 20:21sensor_dirtycomment_update
08-14 19:21sensor_dirtyengagement_update
08-14 18:40repriceThe reobservation adds only minor engagement and no independent deployment, benchmark, or reliability evidence. The case remains an implementation announcement awaiting practical validation.
08-14 18:40alert_silentNothing consequential changed beyond a small score increase; the next briefing is sufficient unless an independent heterogeneous-machine deployment or measured throughput result appears.
08-14 18:40alert_routeNothing consequential changed beyond a small score increase; the next briefing is sufficient unless an independent heterogeneous-machine deployment or measured throughput result appears.
08-14 18:37alert_silentThe repository and Show HN post establish that Lumabri exists and proposes P2P MoE expert execution, but they add no independent throughput, heterogeneous-peer reliability, or practical deployment res
08-14 18:37surface_candidateThe repository and Show HN post establish that Lumabri exists and proposes P2P MoE expert execution, but they add no independent throughput, heterogeneous-peer reliability, or practical deployment res
08-14 18:37alert_routeThe repository and Show HN post establish that Lumabri exists and proposes P2P MoE expert execution, but they add no independent throughput, heterogeneous-peer reliability, or practical deployment res
08-14 18:31groundThe radar already tracks this exact unresolved development in `radar:lumabri-peer-to-peer-moe-inference`, including the need for independent throughput and reliability testing. It still bears directly
08-14 18:28promote_anchororigin walk conf 0.98
08-14 18:27createThe first-party repository is a concrete implementation of a distinct peer-to-peer architecture for distributing sparse-model inference.