Colibrì can run GLM-5.2 locally, a model with approximately 744 billion parameters, using the CPU and loading the components it needs from an SSD. A dedicated graphics card is optional. The documented setup does, however, use approximately 429 GB of model files, and response time depends on available memory, SSD access and CPU computation.

For an entrepreneur, the useful question is which activity the system can handle on an existing computer. The analysis below uses the specifications of an ASUS Zenbook Duo UX8406MA: Core Ultra 9 185H, 32 GB RAM, Windows 11 Pro and an approximately 2 TB NVMe SSD. The hardware inventory was read on 16 November 2026. GLM-5.2 performance on this laptop remains to be measured; the speed figures presented separately come from the project's documentation.

What the 744 billion parameters represent

GLM-5.2 is the pretrained model developed by Z.ai. Colibrì is the inference engine that performs its computations. Installing them on a laptop enables the use of an existing model; its training took place earlier. Customizing an application for a company's documents is a separate step from training the foundation model.

GLM-5.2 uses a Mixture of Experts architecture, abbreviated to MoE. At each step, a routing mechanism selects a subset of experts. The published configuration specifies 256 routed experts and the selection of eight for each token, in the relevant layers, alongside shared components. The project describes approximately 40 billion active parameters per token out of the 744 billion total. A token is a unit of text that may represent part of a word, a word or a symbol.

Selecting experts reduces the computation required for a step compared with activating every expert. The files for all experts remain necessary because other requests may activate different combinations. This creates the distinction between memory used during execution and space occupied on the SSD.

The path of the weights: SSD → RAM → CPU

  1. 01 · STORAGE
    NVMe SSD

    Holds the model files, including experts that are not loaded into working memory.

    429 GB for the reference package

  2. 02 · MEMORY
    RAM and cache

    Contains resident components and a selection of experts. Cache misses require reads from the SSD.

    Budget limited by available memory

  3. 03 · EXECUTION
    CPU

    Performs the operations needed for generation. The test without a GPU explicitly selects the CPU backend.

    Tokens form the response

Why the model fits on an SSD even though it does not fit in RAM

Colibrì keeps experts on disk and brings them into working memory when needed. An expert already available in the cache can be reused without the same SSD read. When requests activate different experts, the engine loads other data. A response obtained by repeating the same question must be distinguished from a conversation with varied questions.

The verified reference package is GLM-5.2 in int4 format with groups of 64 and an int8 MTP component. The published file sizes total 429,276,220,139 bytes: approximately 429.3 GB or 399.8 GiB. The difference between GB and GiB explains some of the apparently contradictory values displayed by operating systems and download pages.

Quantization reduces the numerical precision and size of the weights. Choosing a different format changes the comparison: a smaller file with more aggressive compression must be evaluated as a separate variant. Moving the same weights between SSD and RAM concerns data placement; quantization concerns their numerical representation. The total parameter count alone does not describe response quality.

What the verified ASUS Zenbook Duo configuration supports

Local inventory from 16 November 2026
ComponentObserved valuePractical implication
CPUIntel Core Ultra 9 185H; 16 cores, 22 logical processorsA configuration for a CPU test; the optimal thread count is established through measurement.
Memory32 GB installed; approximately 9.3 GiB available at inventory timeMemory must be freed and RAM pressure checked before testing.
GraphicsIntegrated Intel ArcThe CPU test excludes graphics acceleration. Integrated graphics uses shared memory.
SSDWD PC SN740 NVMe, approximately 2 TBCapacity is known; throughput for expert access requires a separate test.
Free spaceC: approximately 143 GB; F: approximately 832 GBThe 429 GB package fits on F:. C: has insufficient space for it.
SystemWindows 11 Pro, 64-bitNative Windows execution is the initial choice for this configuration.

C: and F: are partitions of the same physical SSD. Putting the model on F: addresses space allocation but does not create a second device that operates independently. Windows activity and access to the model use the same SSD.

After a single copy of the large package, approximately 403 GB of the roughly 832 GB free on F: would remain, before other files and operations. The plan keeps one copy of the model. Free memory at inventory time reflects the applications open at that moment, rather than the entire installed capacity.

How the free space on F: would be divided

429 GB · GLM-5.2 package

403 GB · remaining space

Projected allocation using the free space read on 16 November 2026, before downloading; values in decimal GB.

What the Colibrì project's published measurements show

The project's documentation records a reference setup with WSL2, 12 cores, 25 GB RAM and NVMe access through a VHDX virtual disk. For generation with a cold cache, the published range is approximately 0.05–0.1 tokens per second. These figures belong to the system described by the developers and are not measurements on the Zenbook.

At that rate, generating 200 tokens would take approximately 33–67 minutes by arithmetic. The calculation covers generation only; processing the request and any reasoning tokens add time. Engine startup, the first token and completion of the response are three different moments that must be measured separately.

The laptop's SSD capacity cannot be converted into a speed estimate. Accessing experts differs from copying a large file, and read traffic depends on the cache and request pattern. A useful evaluation records the engine version, model format and cache state alongside every result.

The practical comparison: an 8B model and GLM-5.2

For the first test, we select Qwen3-8B in the GGUF Q4_K_M variant published by the Qwen team. The file is approximately 5.03 GB and can fit entirely in RAM on this configuration, with additional memory for the context and engine. It is an accessible reference for evaluation, without presenting it as the highest-performing model available.

The small model can run with llama.cpp, while GLM-5.2 runs with Colibrì. This compares two complete solutions for the same activity. Differences in the results may come from the model, quantization and engine, so the test does not isolate the effect of parameter count. Both tests initially use the CPU and the same machine.

Proposed configurations; performance results have not been measured
CriterionQwen3-8B Q4_K_MGLM-5.2 int4 gs64 + MTP int8
Enginellama.cpp, CPUColibrì, CPU
Model filesApproximately 5.03 GBApproximately 429.3 GB
Weight placementCan fit entirely in RAMExperts on SSD, limited RAM cache
First evaluation questionDoes it correctly handle routine business tasks?Does it deliver a quality advantage that justifies the measured time?
Speed on this ASUSTo be measuredTo be measured

Installation and testing order on Windows

We start with the small model, a single session and a context of 4,096 tokens chosen for the test. We pin the llama.cpp version and Q4_K_M file, use the local model path and disable GPU offloading. After freeing memory, we record actual consumption, time to the first token and the duration of the complete response.

The large-model experiment uses a separate Colibrì installation, starting from the documented v2.1.0 release and the precisely identified GLM-5.2 package. The model directory is selected on F: before downloading. Setup explicitly selects the model and CPU backend: the automatic menu may recommend a smaller model depending on the resources it detects.

Before starting, we check the configuration recognized by the engine and the placement of the data. The first GLM test is limited to 32 new tokens and 15 minutes of execution, as the budget selected for this experiment. A run stopped at the limit is recorded as such. Longer quality tests follow if the short test's duration is acceptable for the intended activity.

Context and reasoning mode are recorded for each model. The generation token limit includes any reasoning, according to the engine's behavior. For offline use, the files must already be local, and requests to optional services, search or other machines are excluded from that test. The behavior is confirmed by a run with external access disabled.

A woman works at a laptop in an office
Mikhail Nilov

How we evaluate usefulness for a business

We use demonstration requests with verifiable answers: extracting products from an enquiry, identifying stock below a given threshold, summarizing a message and drafting a business reply based exclusively on the supplied terms. Both models receive the same input information. The demonstration data does not come from clients.

For example, given stock levels A = 12, B = 3 and C = 0, the request “identify products with fewer than four units” must return B and C with the correct values. A fluent response that changes the threshold or invents restocking dates is wrong for this task. This separates factual correctness from the impression created by the wording.

From a technical test to a working decision

  1. FIXED REQUEST
    The same information

    A task set and expected answers established before execution.

  2. MEASUREMENT
    Time and resources

    First token, complete duration, memory and disk access, with cache state specified.

  3. REVIEW
    A usable result

    Correct facts, respected conditions and recorded human corrections.

We record the first request after startup, its repetition and a set of different questions separately. A new process may still benefit from the Windows cache; we call it a “cold cache” only when the procedure used justifies that label. For repeated runs, we report the median and variation, along with errors or interruptions.

For interactive work, response time matters directly. For background processing, the number of documents that can be completed within the available working interval also matters. In both situations, selection starts from accepted answers and total time, including human review and correction. A large model becomes a useful option when its measured advantage matches the company's task.

WebHat uses a customized local AI system and provides consultancy on choosing a setup and adapting AI to business workflows. Through custom applications with AI integration, the selected model can be connected to document workflows or internal operations, with requests and results defined for that activity. In this guide, the laptop configuration has been inventoried and the comparison sheet is prepared for measurement.