BSD Ports

misc/colibri g20260717 FreeBSD

Run GLM-5.2 (744B MoE) on a 25GB-RAM consumer machine

System
FreeBSD
Repository
FreeBSD ports
Maintainer
[email protected]
Port added
Last update
Commit
cf6b05e65b
License
APACHE20
Categories
misc

Description

colibri allows to run GLM-5.2 (744B-parameter MoE) on a consumer machine with ~25 GB of RAM - in pure C, with zero dependencies, by streaming experts from disk.

Colibri is a lightweight, quality-preserving MoE runtime that treats VRAM, RAM, and storage as one managed memory hierarchy. Insufficient fast memory may reduce speed, but the default policy never silently changes model precision or router semantics.

Install

Binary package:

pkg install colibri-ai

From the ports tree:

cd /usr/ports/misc/colibri/ && make install clean

To depend on this port in another port's Makefile:

RUN_DEPENDS=	colibri-ai>0:misc/colibri

Packages

The version each package repository carries. “-” means the repository builds no package of it, “n/a” that there is no such repository.

Repositoryaarch64amd64armv6armv7i386
FreeBSD:14:latest-g20260717---
FreeBSD:14:quarterly-g20260717n/a--
FreeBSD:15:latest-g20260717n/a--
FreeBSD:15:quarterly-g20260717n/an/an/a
FreeBSD:16:latest-g20260717n/an/an/a
FreeBSD:16:quarterlyn/an/an/an/an/a

Dependencies

As this port's Makefile names them.

Runtime 8

The full lists below include those, and their dependencies in turn.

Build dependencies 8
Run dependencies 482

The first 100 are listed.

No port depends on this one.

Configuration

No options to configure.

USES
gmake python shebangfix
Service scripts
colibri

Files

pkg-plist 22
  • bin/coli
  • libexec/colibri/doctor.py
  • libexec/colibri/glm
  • libexec/colibri/olmoe
  • libexec/colibri/openai_server.py
  • libexec/colibri/resource_plan.py
  • libexec/colibri/tools/__init__.py
  • libexec/colibri/tools/benchmark_cuda_fixture.py
  • libexec/colibri/tools/clean.py
  • libexec/colibri/tools/convert_fp8_to_int4.py
  • libexec/colibri/tools/convert_olmoe.py
  • libexec/colibri/tools/download_glm52.py
  • libexec/colibri/tools/eval_glm.py
  • libexec/colibri/tools/fetch_benchmarks.py
  • libexec/colibri/tools/gen_unicode.py
  • libexec/colibri/tools/glm_fp8_emit.py
  • libexec/colibri/tools/make_glm_bench_model.py
  • libexec/colibri/tools/make_glm_oracle.py
  • libexec/colibri/tools/quant_ablation.py
  • libexec/colibri/tools/route_coupling_report.py
  • libexec/colibri/tools/route_pairs.py
  • libexec/colibri/tools/run_tests.py
distinfo 1
FileSizeSHA256
JustVugg-colibri-g20260717-8a1dcca_GH0.tar.gz3.055854e+0678a4899981bd2da030b48efa328c3c6b330252b08425d154f58acfbbd23edf6a

Commit history 4

DateCommitCreditsLog message
cf6b05e65b Yuri Victorovich misc/colibri: update g20260715 → g20260717
Also: * add service * upgrade optimization O2->O3 * add Python dependencies
df28e83ded Yuri Victorovich misc/colibri: update g20260714 → g20260715
1fa4edff26 Yuri Victorovich misc/colibri: Minor change in the comment
3dd6a7a2a9 Yuri Victorovich misc/colibri: New port: Run GLM-5.2 (744B MoE) on a 25GB-RAM consumer machine