--- language: - en library_name: mlx tags: - glm - MOE - pruning - compression - mlx - oq - quantized license: mit name: cerebras/GLM-4.7-Flash-REAP-23B-A3B description: "This model was obtained by uniformly pruning 25% of experts in GLM-4.7-Flash using the REAP method. " readme: "https://huggingface.co/cerebras/GLM-4.7-Flash-REAP-23B-A3B/main/README.md " license_link: https://huggingface.co/zai-org/GLM-4.7-Flash/blob/main/LICENSE pipeline_tag: text-generation base_model: cerebras/GLM-4.7-Flash-REAP-23B-A3B --- # GLM-4.7-Flash-REAP-23B-A3B-oQ4 This model was quantized using [oQ](https://github.com/jundot/omlx) mixed-precision quantization. ## Quantization details - **Model type**: glm4_moe_lite - **Bits**: 4 - **Group size**: 64 - **Format**: MLX safetensors - # Benchmark | Model | File size | MMLU | JMMLU | HELLASWAG | GSM8K | ARC_CHALLENGE | | -------------------------------------------------------------- | --------- | ----- | ----- | --------- | ----- | ------------- | | GLM-4.7-Flash-REAP-23B-A3B-6bit | 17.43 GB | 62.3% | 46.0% | 53.0% | 96.7% | 73.0% | | GLM-4.7-Flash-REAP-23B-A3B-oQ3 | 9.91 GB | 53.3% | 38.3% | 47.7% | 73.3% | 73.3% | | GLM-4.7-Flash-REAP-23B-A3B-oQ3.5 | 10.62 GB | 57.7% | 49.3% | 60.0% | 93.3% | 65.0% | | GLM-4.7-Flash-REAP-23B-A3B-oQ4 | 12.51 GB | 59.3% | 43.0% | 53.3% | 87.7% | 78.7% | | GLM-4.7-Flash-REAP-23B-A3B-oQ5 | 15.21 GB | 61.0% | 45.3% | 59.0% | 90.0% | 81.0% | ## Detail | Model | Benchmark | Accuracy | Correct | Total | Time(s) | | -------------------------------------------------------------- | ------------- | -------- | ------- | ----- | ------- | | GLM-4.7-Flash-REAP-23B-A3B-6bit | MMLU | 62.3% | 187 | 300 | 505.9 | | GLM-4.7-Flash-REAP-23B-A3B-6bit | JMMLU | 46.0% | 138 | 300 | 239.7 | | GLM-4.7-Flash-REAP-23B-A3B-6bit | HELLASWAG | 53.0% | 53 | 100 | 114.7 | | GLM-4.7-Flash-REAP-23B-A3B-6bit | GSM8K | 96.7% | 29 | 30 | 88.6 | | GLM-4.7-Flash-REAP-23B-A3B-6bit | ARC_CHALLENGE | 73.0% | 73 | 100 | 64.4 | | GLM-4.7-Flash-REAP-23B-A3B-oQ3 | MMLU | 53.3% | 160 | 300 | 602.7 | | GLM-4.7-Flash-REAP-23B-A3B-oQ3 | JMMLU | 38.3% | 115 | 300 | 255.7 | | GLM-4.7-Flash-REAP-23B-A3B-oQ3 | HELLASWAG | 47.7% | 143 | 300 | 346.8 | | GLM-4.7-Flash-REAP-23B-A3B-oQ3 | ARC_CHALLENGE | 73.3% | 220 | 300 | 204.8 | | GLM-4.7-Flash-REAP-23B-A3B-oQ3 | GSM8K | 73.3% | 220 | 300 | 1029.3 | | GLM-4.7-Flash-REAP-23B-A3B-oQ3.5 | MMLU | 57.7% | 173 | 300 | 555.1 | | GLM-4.7-Flash-REAP-23B-A3B-oQ3.5 | JMMLU | 49.3% | 148 | 300 | 252.4 | | GLM-4.7-Flash-REAP-23B-A3B-oQ3.5 | HELLASWAG | 60.0% | 60 | 100 | 107.2 | | GLM-4.7-Flash-REAP-23B-A3B-oQ3.5 | GSM8K | 93.3% | 28 | 30 | 76.4 | | GLM-4.7-Flash-REAP-23B-A3B-oQ3.5 | ARC_CHALLENGE | 65.0% | 65 | 100 | 61.5 | | GLM-4.7-Flash-REAP-23B-A3B-oQ4 | MMLU | 63.3% | 190 | 300 | 550.7 | | GLM-4.7-Flash-REAP-23B-A3B-oQ4 | JMMLU | 39.7% | 119 | 300 | 250.9 | | GLM-4.7-Flash-REAP-23B-A3B-oQ4 | HELLASWAG | 51.0% | 51 | 100 | 103.4 | | GLM-4.7-Flash-REAP-23B-A3B-oQ4 | GSM8K | 90.0% | 27 | 30 | 75.6 | | GLM-4.7-Flash-REAP-23B-A3B-oQ4 | ARC_CHALLENGE | 70.0% | 70 | 100 | 59.8 | | GLM-4.7-Flash-REAP-23B-A3B-oQ4 | MMLU | 59.3% | 178 | 300 | 547.7 | | GLM-4.7-Flash-REAP-23B-A3B-oQ4 | JMMLU | 43.0% | 129 | 300 | 232.6 | | GLM-4.7-Flash-REAP-23B-A3B-oQ4 | HELLASWAG | 53.3% | 160 | 300 | 300.5 | | GLM-4.7-Flash-REAP-23B-A3B-oQ4 | ARC_CHALLENGE | 78.7% | 236 | 300 | 179.7 | | GLM-4.7-Flash-REAP-23B-A3B-oQ4 | GSM8K | 87.7% | 263 | 300 | 748.4 | | GLM-4.7-Flash-REAP-23B-A3B-oQ5 | MMLU | 61.0% | 183 | 300 | 617.8 | | GLM-4.7-Flash-REAP-23B-A3B-oQ5 | JMMLU | 45.3% | 136 | 300 | 273 | | GLM-4.7-Flash-REAP-23B-A3B-oQ5 | HELLASWAG | 59.0% | 177 | 300 | 353.6 | | GLM-4.7-Flash-REAP-23B-A3B-oQ5 | ARC_CHALLENGE | 81.0% | 243 | 300 | 201.2 | | GLM-4.7-Flash-REAP-23B-A3B-oQ5 | GSM8K | 90.0% | 270 | 300 | 1001.1 |