SaaS Metrics
The core product is 'colibri', an engine designed to run large Mixture-of-Experts (MoE) models (GLM-5.2, 744B MoE) on consumer-grade hardware. Its key architectural feature is streaming experts from disk for CPU-only execution to manage RAM constraints. The discussion focuses on benchmarking its performance.
Raw Developer Origin & Technical Request
GitHub Issue
Jul 10, 2026
## Commit
`3e4d08b6bfe67b9bc17e8a0694b7c2604862612d`
## Hardware and storage
- CPU: AMD Ryzen 9 9950X 16-Core Processor
- Threads: 32 logical CPUs
- RAM: 128 GiB total
- Model storage: Samsung SSD 9100 PRO 1TB NVMe
- Storage link: PCIe 5.0 x4
- Filesystem: local ext4
## Software environment
- Host OS: Linux
- Build/runtime path: CPU-only default build, no CUDA
- Model location: `/models/glm52-colibri-int4`
- Engine: `GLM-5.2 · 744B MoE · int4 · streaming CPU`
## Build and benchmark commands
```sh
cd /home/colibri/c
./setup.sh
./coli info --model /models/glm52-colibri-int4
./coli run --model /models/glm52-colibri-int4 --ngen 64 --temp 0 "Explain in one short paragraph how a CPU-only streaming MoE model trades RAM, disk, and compute."
./coli run --model /models/glm52-colibri-int4 --ngen 64 --temp 0 "Explain in one short paragraph how a CPU-only streaming MoE model trades RAM, disk, and compute."
```
Disk throughput check on a representative shard:
```sh
gcc -O2 -fopenmp c/iobench.c -o /tmp/colibri-iobench
/tmp/colibri-iobench /models/glm52-colibri-int4/out-00000.safetensors 19 64 8 1
```
## Results
Warm-up policy: two back-to-back generation runs after relocating the model to the PCIe 5.0 drive. The learned expert-usage history was retained.
Run count: 2
Generation throughput:
- Run 1: `64 token in 228.22s (0.28 tok/s)`
- Run 2: `64 token in 230.59s (0.28 tok/s)`
- Median throughput: `0.28 tok/s`
Run 1 profile:
- RSS: `82.15 GB`
- Expert hit-rate: `56.7%`
- Profi...
Developer Debate & Comments
No active discussions extracted for this entry yet.
Adjacent Repository Pain Points
Other highly discussed features and pain points extracted from JustVugg/colibri.
stats
Frequently Asked Questions
Market intelligence mapped to The core product is 'colibri', an engine designed to run large Mixture-of-Experts (MoE) models (GLM-5.2, 744B MoE) on consumer-grade hardware. Its key architectural feature is streaming experts from disk for CPU-only execution to manage RAM constraints. The discussion focuses on benchmarking its performance..
What is the technical positioning of The core product is 'colibri', an engine designed to run large Mixture-of-Experts (MoE) models (GLM-5.2, 744B MoE) on consumer-grade hardware. Its key architectural feature is streaming experts from disk for CPU-only execution to manage RAM constraints. The discussion focuses on benchmarking its performance.?
Are engineers actively discussing The core product is 'colibri', an engine designed to run large Mixture-of-Experts (MoE) models (GLM-5.2, 744B MoE) on consumer-grade hardware. Its key architectural feature is streaming experts from disk for CPU-only execution to manage RAM constraints. The discussion focuses on benchmarking its performance.?
What architecture is tied to The core product is 'colibri', an engine designed to run large Mixture-of-Experts (MoE) models (GLM-5.2, 744B MoE) on consumer-grade hardware. Its key architectural feature is streaming experts from disk for CPU-only execution to manage RAM constraints. The discussion focuses on benchmarking its performance.?
Engagement Signals
Cross-Market Term Frequency
Quantifies the cross-market adoption of foundational terms like RSS and tok/s by tracking occurrence frequency across active SaaS architectures and enterprise developer debates.