Blog
Runtime Deep DiveColibrì v1.4.014 min readAugust 1, 2026

Colibri Can Run GLM-5.2 on 25GB RAM. Here Is the Catch.

A tiny open inference engine turns NVMe storage into a memory tier for 744B to 2.8T MoE models. The result is technically remarkable and brutally slow at the minimum.

Short answer: yes, Colibrì has demonstrated GLM-5.2 on a machine with 25GB of RAM by keeping the dense core resident and streaming routed experts from NVMe. But the model still occupies about 372GB on disk, and the project reports only 0.05 to 0.1 tokens per second on its 25GB baseline. This is a research breakthrough, not a fast laptop deployment.

25GB
Proven RAM floor
372GB
GLM container
9.9GB
Dense resident core
4
Model families

What Colibri is, and what it is not

Colibrì is an Apache 2.0 inference engine and research platform written primarily in C. It is not a model, a new quantization of GLM-5.2, or a smaller distilled checkpoint. Its central idea is to treat VRAM, system RAM and fast storage as one managed hierarchy instead of requiring every parameter to remain in expensive memory.

The engine currently supports four distinct MoE families through sibling runtime implementations: GLM-5.2, Thinking Machines' Inkling, Moonshot AI's Kimi K3 and AI2's OLMoE. The same command-line, server and web interfaces sit above them.

How can a 744B model use only 9.9GB of resident RAM?

The answer is sparsity. GLM-5.2 contains 744B total parameters, but only about 40B are active for each token. Colibrì separates the weights that must remain available on every pass from the routed experts that change token by token.

Fast tier

VRAM

Hot experts and GPU work when available

Resident tier

System RAM

Dense attention, shared experts and a learned cache

Capacity tier

NVMe storage

About 370GB of GLM-5.2 routed experts, loaded on demand

The much-shared 9.9GB resident number describes the dense part of the int4 GLM-5.2 container. It does not mean GLM-5.2 has become a 9.9GB model. The routed expert bank still exists, still needs hundreds of gigabytes of storage and still has to cross the storage bus when the router selects an expert that is not cached.

The measured speed tells the real story

Colibrì's documentation is unusually clear that there is no speed SLA. The following figures are project-reported measurements on different community and development systems, not LocalClaw benchmark reproductions. They show how the same model moves from proof-of-possibility to something closer to interactive use as more weights become resident.

Hardware shapeReported decodeWhat it means
25GB development box0.05-0.1 tok/sTen to twenty seconds per token. A correctness and accessibility proof.
Single RTX 5070 Ti laptop-class system1.07 tok/sUsable for patient experimentation, still slow for daily chat.
128GB CPU-only desktop~1.8 tok/s warmA more credible private research workstation path.
6 x RTX 5090, full expert residency5.8-6.8 tok/sInteractive, but no longer ordinary consumer hardware.

Storage speed, cache state, prompt routing, backend and expert residency all change performance. A second independent NVMe can help because Colibrì can distribute expert reads across drives, but the project itself asks users to measure rather than assume a universal gain.

Which models does Colibri run?

FamilyScaleStorage / resident realityPractical verdict
GLM-5.2744B / 40B active372GB int4 container; 9.9GB dense resident coreBest documented Colibrì path
Inkling975B / 41B activeAbout 469GB; optional 15.3GB quantized dense setRunnable on 25GB, very slow from NVMe
Kimi K32.8T / 104B activeAbout 1.45TB of routed experts; ~35GB dense set at 4-bitText-only research path, not a daily driver
OLMoE7B / 1B activeConverted local containerSmall validation and experimentation target

Engine support does not change each model's license. Colibrì itself is Apache 2.0, while Kimi K3 remains governed by Moonshot's custom model license and other weights retain their own terms.

Does this make Kimi K3 a consumer local model?

No. It makes a previously unthinkable text-only research run technically possible. Colibrì's Kimi K3 implementation streams the checkpoint's native MXFP4 experts directly and quantizes the dense set at load time. The project documents approximately 1.45TB of expert data and a resident dense set of about 35GB at 4-bit.

In a measured 93-layer run, the project reports that direct I/O reduced output time from roughly 21 seconds per token to about 9.4 seconds per token. That is meaningful systems progress, but it remains far from an interactive assistant. The Colibrì path is also text-only and does not load Kimi K3's vision tower.

Installation and platform support

Colibrì publishes prebuilt archives for Linux, macOS and Windows, while the engine can also be built from source with GCC or Clang plus OpenMP. The engine is pure C, but the launcher, conversion helpers and API gateway use Python 3. The model download is the real installation cost: GLM-5.2 alone needs about 372GB plus safe free-space headroom.

Use v1.4.0 or newer for the new model families

The August 1 release notes say earlier v1.3 archives omitted the extra model binaries, so Kimi K3 and Inkling could not work from those packaged downloads. Version 1.4.0 corrects that packaging issue. This is a fast-moving project: pin the exact release, container and command when reporting a result.

Why this matters for LocalClaw and OpenClaw

Colibrì exposes an OpenAI-compatible API with /v1/chat/completions, plus an Anthropic-compatible /v1/messages endpoint. That makes it conceptually relevant to local control centers and agent tools: a compatible client can point at the local server instead of a hosted provider.

But compatibility is not the same as product readiness. LocalClaw should treat Colibrì as an experimental advanced runtime, not silently install it or recommend GLM-5.2 to ordinary Macs. A future integration would need reproducible macOS tests, disk-space checks, a clear speed warning, version pinning and a recovery path before it belongs in a one-click setup.

Use it now if

You are doing inference research

You have hundreds of gigabytes of fast storage, accept experimental setup work and want to study enormous MoE models locally.

Skip it for now if

You want a fast daily assistant

A smaller GGUF or MLX model that fits in RAM will be dramatically simpler, faster and easier to maintain.

LocalClaw verdict

Colibrì is one of the most interesting local inference projects of 2026 because it attacks the memory wall instead of merely publishing another quant. It proves that a model can remain semantically intact while its weights are staged across storage, RAM and VRAM according to measured routing behavior.

The honest conclusion is equally important: runnable does not mean comfortable. On 25GB, GLM-5.2 is a slow experiment. At 128GB it becomes more credible. With six RTX 5090 cards and full residency it becomes interactive. Colibrì expands what is possible on hardware you own, but it does not repeal bandwidth, capacity or latency.

Frequently asked questions

Is Colibri an open-source model?

No. Colibri is an Apache 2.0 inference engine and research platform. It runs supported model families whose weights retain their own licenses.

Can GLM-5.2 really run with 25GB of RAM?

Yes, the project reports a proven 25GB baseline by streaming experts from a roughly 372GB NVMe container. Performance at that floor is only 0.05 to 0.1 tokens per second.

Does 9.9GB resident mean GLM-5.2 is a 9.9GB model?

No. That figure covers the dense resident part of the int4 runtime. Roughly 370GB of routed experts remain on storage and are loaded or cached as needed.

Can Colibri run Kimi K3 on a normal PC?

It can technically stream the text model from about 1.45TB of expert data, but the documented output remains measured in seconds per token. This is a research path, not normal interactive use.

Is Colibri compatible with OpenAI clients?

Yes. Its server implements OpenAI-compatible model and completion endpoints and also exposes an Anthropic Messages API endpoint. Client compatibility does not guarantee practical speed on minimum hardware.

Sources