--- base_model: - zai-org/GLM-5.2 license: mit language: - en - zh pipeline_tag: text-generation tags: - glm_moe_dsa --- ## Overview * **Original Model:** [zai-org/GLM-5.2](https://huggingface.co/zai-org/GLM-5.2) This repository contains the **GGUF** quantized files for **[zai-org/GLM-5.2](https://huggingface.co/zai-org/GLM-5.2)**. | Version | FFN Exps Gate-Up-Down | Token Embedding | Output Weight | Size | | :--- | :--- | :--- | :--- | :--- | | IQ2_M | IQ2_XXS / IQ2_XXS / IQ3_S | Q8_0 | Q8_0 | 262 GB | | MixQ5-IQ3_M | IQ3_S / IQ3_S / Q5_K | BF16 | BF16 | 431 GB | | MXFP4 | MXFP4 / MXFP4 / MXFP4 | BF16 | BF16 | 443 GB | **Note:** The frontmost FFN layers in the first block are all set to Q8_0, while the remaining tensors remain in BF16 and FP32 formats. ### The "Garbage In, Garbage Out" Hypothesis > Please note that the Quantization Override (Q8) applied to the very first layers in these custom models is an intentional experiment. The core hypothesis we are testing challenges the "Garbage In, Garbage Out" concept: if the frontmost layers (which capture and process the initial input context) are extensively compressed, will it inevitably trap the model with degraded inputs and result in poor final generations? > > And conversely: if the frontmost layers are aggressively preserved, will it allow the model to capture clear and clean inputs, leading to superior final generation results? > > **Call for Feedback:** If you are willing to test these models, please share your performance benchmarks, perplexity scores, or qualitative observations in the Community Discussions or Issues tab. Your field reports are crucial to help determine whether early-layer degradation is a fatal flaw or a manageable trade-off in MoE architectures. --- **Prompt used:** > "Write a single HTML file with a full-page canvas and no libraries. Simulate a realistic side-view of a moving car as the main subject. Keep the car visible in the foreground while the background landscape scrolls continuously to create the feeling that the car is driving forward. Use layered scenery for depth: nearby ground, roadside elements, trees, poles, and distant hills or mountains should move at different speeds for a natural parallax effect. Animate the wheels spinning realistically and add subtle body motion so the car feels connected to the road. Let the environment pass smoothly behind it, with repeating but varied scenery that makes the movement feel believable. Use cinematic lighting and a cohesive sky, such as sunset, dusk, or daylight, to enhance atmosphere. The overall motion should feel calm, immersive, and realistic, with a seamless looping animation." --- ### IQ2_M-Max-Result Here are the results for IQ2_M on 4 x RTX Pro 6000 : * Thinking Effort: Max * Reasoning Token : 103.879 * Output : 10.386 Thinking Max is truly absurd, it spent 2,417 seconds of reasoning on 4 RTX Pro 6000s and burned approximately 100,000 tokens just for its internal thinking. The output is as shown in the preview here. **Download :** GLM-5.2-IQ2_M 
