Price terms as published
Each line is the vendor’s own wording, split from the price capture checked on 2026-09-22. Nothing here is rounded, converted or re-worded — the full sentences stay in the review below. Lines that carry only a price continue the plan named on the line above, in the vendor’s own order.
| # | Plan / term | Checked |
|---|---|---|
| 1 | Free (open source, Apache-2.0) | 2026-09-22 |
One-liner: Colibri (Italian for hummingbird) is an open-source inference engine written in pure C with zero runtime dependencies that runs frontier mixture-of-experts models — from 744B up to 2.8T parameters — on consumer hardware by treating storage, RAM, and VRAM as a single memory hierarchy and streaming routed experts from disk on demand.
Verified facts (GitHub API, 2026-09-22 00:26 UTC+8)
| Attribute | Value | Source |
|---|---|---|
| GitHub stars | 36,752 (3,927 forks) | JustVugg/colibri via GitHub API, 2026-09-22 |
| License | Apache-2.0 | GitHub API license field |
| Language | C | GitHub API |
| Created | 2026-07-01 | GitHub API |
| Last push | 2026-09-21 — active daily | GitHub API |
| Latest release | v1.12.0 | repo README (releases badge + CLI banner) |
| Model roster (README) | GLM-5.2/5.3 (744B), GLM-5.3-Flash (321B, vision), Inkling (975B), Kimi K3 (2.8T), DeepSeek V4 Flash (284B), DeepSeek V4.1 Flash (552B, vision), Qwen3.8-Flash-Next (125B), Qwen3.6 (35B-A3B), OLMoE (7B) | repo README, “nine families run today” |
| Signature demo | GLM-5.2 744B int4 resident in 9.9 GB, answers on a ~25 GB machine | README banner output |
| Front end | coli chat / coli serve / coli web (dashboard) |
README |
Star-growth corroboration (secondary sources)
- 07-09 Show HN: ~1,600 stars in a day, 453–769 points (ai-tldr.dev, flowtivity.ai).
- ~17,000 stars three weeks after launch (YouTube explainer, early Aug).
- 26–27k on several trackers as of 2026-09-08; a single tracker showed ~36k that same day — so 36.8k on 09-22 means most of September’s growth is real and the “stuck at 27k” snapshot is what went stale. All figures are time-stamped secondary readings; the API row above is the primary one.
How it works (and why your laptop can hold a 744B model)
An MoE only activates a small slice of its parameters per token. Colibri keeps
the dense layers resident (~9.9 GB for GLM-5.2) and streams the routed experts
from SSD as the router calls for them — storage I/O, placement, and scheduling
replacing GPU VRAM. No BLAS, no CUDA at runtime, no Python: gcc + OpenMP on
x86-64 with AVX2. A newer CUDA tier and opt-in CPU/GPU overlap levers
(DSV4_HYBRID, COLI_CUDA_MOE_DOUBLE) exist, default off, “byte-identical
when unset”.
The catches
- Speed is the price. The launch-era cold throughput for the flagship 744B demo was on the order of 0.05–0.1 tok/s on a 25 GB laptop (secondary reporting; the README itself makes no speed promise). The project states this explicitly: “no SLA on speed” — insufficient fast memory reduces speed but “must not quietly redefine the model”.
- It is a research platform. Aggressive systems experiments land here; you pin versions, not a support contract.
- One-machine benchmark culture. Numbers are author-posted; community reproduction is still thin for a repo this young.
Who should use it
Local-inference tinkerers who want frontier MoE weights on hardware they own, and anyone prototyping “big model, small box” stacks (retrieval frontends, offline assistants with latency slack). If you need sub-second interactive coding, buy VRAM instead.
Stars, license, language, dates, release, and model roster pulled from the GitHub API and raw README on 2026-09-22 (00:26–00:31 UTC+8). Growth-history figures are time-stamped third-party snapshots, labeled as such. No inference run was executed on our hardware for this page; no hands-on claims are made.