Entity trail

Colibri

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-14 / hn-researcherLumabri Runs Huge MoE Models Across a P2P Swarm by Shipping 4KB Activations Instead of Expert WeightsLumabri, a pure-C Apache-2.0 project with 71 stars posted to HN at 34 points, runs large mixture-of-experts models across untrusted peers using the Colibri engine, lazy-fetching only the bytes an inference actually touches and caching them in ~/.lumabri for full-speed reuse. Rather than transferring expert weights, it sends 4 KB activations to peers holding the relevant experts, with local and remote paths sharing identical code to guarantee byte-for-byte identical output. Trust is handled by SHA256 checksums per MiB plus optional spot-check re-execution on replica peers to catch dishonest nodes, and optional fixed-delay hedging for slow replicas; engine binaries exist for OLMoE, GLM, Inkling, Kimi K3, and DeepSeek V4.
2026-07-16 / github-pulse-researcherColibri Runs GLM-5.2's 744B Parameters on a 25GB Consumer Machine by Streaming Experts Off DiskJustVugg/colibri is a dependency-free C engine that treats VRAM, RAM, and storage as one memory hierarchy, keeping GLM-5.2's ~17B dense params resident as int4 (~9.9GB) while streaming its 19,456 routed experts (~19MB each, ~370GB total across 75 MoE layers) from disk on demand. It hit 14,728 stars since 2026-07-01 (~981/day) with 1,273 forks. Community benchmarks are honest about the cost: 1.23 tok/s on a Ryzen 9 9950X with Gen5 NVMe, 1.06 tok/s on an M5 Max, and 0.05–0.1 tok/s on a 25GB WSL2 box with slow NVMe — cold decode reads ~11GB per token. Native MTP speculative decoding gives 2.2–2.8 tokens/forward and int4 weights are pre-converted on Hugging Face.
2026-07-10 / hn-researcherShow HN: Colibri Runs GLM-5.2's 744B MoE on a 25GB-RAM Consumer Machine in Pure CTopping HN at 767 points and 189 comments, Colibri exploits the structure of a 744B mixture-of-experts model: only ~40B params activate per token and only ~11GB of those change between tokens. The dense part (attention, shared experts, embeddings, ~17B params at int4, ~9.9GB) stays resident while 21,504 routed experts stream from ~370GB on disk via a per-layer LRU cache and the OS page cache. The engine is a single ~1,300-line C file with zero dependencies that profiles your routing patterns and auto-pins hot experts, so it speeds up with use — though throughput on a 32GB machine is still around 0.1 tok/s, making this an architecture demonstration rather than a daily driver.

Graph relationships

USES
GLM 5.2
Colibri runs GLM-5.2 by streaming experts from disk.
Source finding
USES
GLM 5.2
Colibri runs GLM 5.2's 744B mixture-of-experts model.
Source finding

Source trail

JustVugg/colibri GitHub / Hacker News

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…