{"msg":"OK","data":[{"id":245445,"name":"mlx-community_gemma-4-e4b-it-qat-OptiQ-4bit","nickname":"mlx-community_gemma-4-e4b-it-qat-OptiQ-4bit","description":"mlx-community/gemma-4-e4b-it-qat-OptiQ-4bit 是一款面向苹果芯片本地运行的多模态文本生成模型，基于谷歌经量化感知训练（QAT）的 Gemma-4 基座，采用混合精度量化方案，通过敏感度引导的逐层比特分配，将敏感层保留为 8 位、稳健层压缩至 4 位，在保持约 5.17 比特/权重（磁盘占用约 7.0 GB）的前提下，相较同基座的均匀 4 位量化版本在六项基准测试（MMLU、GSM8K、IFEval、BFCL、HumanEval、HashHop）的平均能力得分上提升 1.19 分，其中 HumanEval 与长上下文任务增益最为显著。该模型支持纯文本生成以及图像与文本联合输入，视觉部分以 bf16 旁路权重单独存放，可兼容标准 MLX 推理流程与统一模型接口，并支持通过草稿模型加速推理。其差异化优势在于将量化感知训练与敏感度驱动的混合精度分配相结合，在有限存储预算下优先保障关键层精度，从而在代码生成、指令遵循和长上下文理解等场景中取得优于均匀低比特量化的表现，适用于需要在苹果芯片设备上本地部署多模态大模型、并兼顾生成质量与资源效率的应用场景。","likes":0,"downloads":0,"path":"mlx-community/mlx-community_gemma-4-e4b-it-qat-OptiQ-4bit","repository_id":330232,"private":false,"tags":[{"name":"text-generation","category":"task","group":"natural_language_processing","built_in":true,"show_name":"text generation","i18n_key":"text-generation","created_at":"2023-12-28T11:58:32.245769Z","updated_at":"2023-12-28T11:58:32.245769Z"},{"name":"apache-2.0","category":"license","group":"","built_in":true,"show_name":"Apache-2.0","i18n_key":"apache-2.0","created_at":"2023-12-28T11:58:32.245769Z","updated_at":"2023-12-28T11:58:32.245769Z"},{"name":"safetensors","category":"framework","group":"","built_in":true,"show_name":"SafeTensors","i18n_key":"safetensors","created_at":"2023-12-28T11:58:32.245769Z","updated_at":"2023-12-28T11:58:32.245769Z"},{"name":"image-text-to-text","category":"task","group":"multimodal","built_in":true,"show_name":"image text to text","i18n_key":"image-text-to-text","created_at":"2024-03-12T04:36:03.249282Z","updated_at":"2025-05-20T02:55:00.120138Z"},{"name":"quantized","category":"task","group":"","built_in":false,"show_name":"","created_at":"2024-05-16T06:18:24.3723Z","updated_at":"2024-05-16T06:18:24.3723Z"},{"name":"Information Technology (IT)","category":"industry","group":"","built_in":true,"show_name":"information technology (it)","i18n_key":"Information Technology (IT)","created_at":"2024-05-16T02:42:12.939Z","updated_at":"2024-05-16T02:42:12.939Z"},{"name":"mlx","category":"task","group":"","built_in":false,"show_name":"","created_at":"2024-06-29T11:26:08.546088Z","updated_at":"2024-06-29T11:26:08.546088Z"},{"name":"4bit","category":"task","group":"","built_in":false,"show_name":"","created_at":"2024-06-29T11:35:15.908314Z","updated_at":"2024-06-29T11:35:15.908314Z"},{"name":"8bit","category":"task","group":"","built_in":false,"show_name":"","created_at":"2024-06-29T11:35:52.431062Z","updated_at":"2024-06-29T11:35:52.431062Z"},{"name":"vllm","category":"runtime_framework","group":"inference","built_in":true,"show_name":"vllm","i18n_key":"Vllm","created_at":"2024-10-18T14:28:26.315003Z","updated_at":"2024-10-18T14:28:26.315003Z"},{"name":"sglang","category":"runtime_framework","group":"inference","built_in":true,"show_name":"sglang","i18n_key":"SGLang","created_at":"2025-01-09T05:43:24.021587Z","updated_at":"2025-01-09T05:43:24.021587Z"},{"name":"evalscope","category":"runtime_framework","group":"evaluation","built_in":true,"show_name":"","created_at":"2025-06-20T13:29:40.619652Z","updated_at":"2025-06-20T13:29:40.619652Z"},{"name":"apple-silicon","category":"task","group":"","built_in":false,"show_name":"","created_at":"2026-06-06T15:23:29.158788Z","updated_at":"2026-06-06T15:23:29.158788Z"},{"name":"gemma-4","category":"task","group":"","built_in":false,"show_name":"","created_at":"2026-06-06T15:23:29.158788Z","updated_at":"2026-06-06T15:23:29.158788Z"},{"name":"mixed-precision","category":"task","group":"","built_in":false,"show_name":"","created_at":"2026-06-06T15:23:29.158788Z","updated_at":"2026-06-06T15:23:29.158788Z"},{"name":"optiq","category":"task","group":"","built_in":false,"show_name":"","created_at":"2026-06-06T15:23:29.158788Z","updated_at":"2026-06-06T15:23:29.158788Z"},{"name":"qat","category":"task","group":"","built_in":false,"show_name":"","created_at":"2026-06-15T22:21:45.066133Z","updated_at":"2026-06-15T22:21:45.066133Z"}],"readme":"","repository":{"http_clone_url":"https://opencsg.com/models/mlx-community/mlx-community_gemma-4-e4b-it-qat-OptiQ-4bit.git","ssh_clone_url":"git@opencsg.com:models/mlx-community/mlx-community_gemma-4-e4b-it-qat-OptiQ-4bit.git"},"default_branch":"","created_at":"2026-09-22T02:31:18.544831Z","updated_at":"2026-10-05T12:48:00.141543Z","widget_type":"","status":"","user_likes":false,"source":"local","sync_status":"completed","enable_inference":false,"enable_finetune":false,"enable_evaluation":false,"disable_inference_reason":"","disable_finetune_reason":"","disable_evaluation_reason":"","base_model":"","license":"apache-2.0","compliance_status":"compliant","commercial_permission":"commercial_allowed","can_write":false,"can_manage":false,"is_organization":false,"namespace":null,"scores":null,"sensitive_check_status":"","mirror_last_updated_at":"0001-01-01T00:00:00Z","metadata":{"model_params":0,"tensor_type":"","architecture":"","mini_gpu_memory_gb":0,"mini_gpu_finetune_gb":0,"model_type":"","class_name":""},"report_url":"","medium_risk_count":0,"high_risk_count":0,"url":"","mirror_task_status":"finished","xnet_enabled":false,"xnet_migration_status":"","xnet_migration_progress":0,"arch_allowed":false,"hf_path":"mlx-community/gemma-4-e4b-it-qat-OptiQ-4bit","ms_path":"","csg_path":""}],"total":207347}