This is an EXL3 quantization of Sage S2.1 Flash, a scientific multimodal model developed by Shanghai Artificial Intelligence Laboratory. The original model is designed for scientific reasoning, coding, information retrieval, and tool use. See its model card for the model's capabilities, evaluation results, and intended use.
| Property | Value |
|---|---|
| Format | EXL3, MCG codebook |
| Quantized tensors | Routed expert gate_proj, up_proj, and down_proj weights |
| Routed expert precision | 4 bits per weight |
| Other tensors | Unquantized; copied from the original Sage S2.1 Flash checkpoint |
| Weights on disk | Approximately 164 GiB, in 83 safetensors shards |
| Configured maximum context | 1,048,576 tokens; practical limits depend on the inference runtime and available memory |
The 4 bpw figure applies to the routed expert weights, not to every tensor or the overall checkpoint. The quantization metadata identifies Sage-S2.1-Flash-FP8 revision 716e6436905650b7be93e1b7dedfb0bb95afd3a3 as the source for the routed experts, and Sage-S2.1-Flash revision 76f4513db7464485605a002a06a6190de0460574 for the unchanged non-routed tensors. See quantization_config.json for the full tensor-level metadata.
pip install -U huggingface_hub
hf download Mia-AiLab/Sage-S2.1-Flash-EXL3-4bpw --local-dir ./sage
The repository includes the tokenizer, chat template, model configuration, shard index, and SHA256SUMS. To verify a complete download:
cd sage
sha256sum -c SHA256SUMS
Use an inference engine that explicitly supports both the Sage S2.1 Flash architecture and this EXL3 routed-expert layout. EXL3 safetensors are quantized weights; a standard Transformers or FP8 loading example for the original model does not establish compatibility with this checkpoint. Memory requirements include the weights, runtime buffers, and context cache, and depend on how the engine distributes or offloads the model.
The configuration and processor files include vision and video settings, but support for those inputs depends on the inference engine. Scientific projectors, encoders, and preprocessing tools are not included in this repository. See the original model card and Sage SciPerception for their requirements; compatibility with this quantization and a particular serving engine must be established separately.
The original model's reported benchmarks are not measurements of this quantization. No independent quality, speed, or multimodal evaluation for this checkpoint is provided here. Quantization can change outputs and accuracy; evaluate it on your own tasks before relying on it. The original model's limitations also apply.
MIT, following the source model's license.
model-00021-of-00083.safetensors 2.1 GBmodel-00045-of-00083.safetensors 2.1 GBmodel-00069-of-00083.safetensors 2.1 GBmodel-00058-of-00083.safetensors 2.1 GBmodel-00023-of-00083.safetensors 2.1 GBmodel-00047-of-00083.safetensors 2.1 GBmodel-00031-of-00083.safetensors 2.1 GBmodel-00051-of-00083.safetensors 2.1 GBmodel-00055-of-00083.safetensors 2.1 GBmodel-00075-of-00083.safetensors 2.1 GBmodel-00036-of-00083.safetensors 2.1 GBmodel-00060-of-00083.safetensors 2.1 GBmodel-00018-of-00083.safetensors 2.1 GBmodel-00029-of-00083.safetensors 2.1 GBmodel-00040-of-00083.safetensors 2.1 GBmodel-00053-of-00083.safetensors 2.1 GBmodel-00066-of-00083.safetensors 2.1 GBmodel-00077-of-00083.safetensors 2.1 GBmodel-00062-of-00083.safetensors 2.1 GBmodel-00038-of-00083.safetensors 2.1 GBmodel-00025-of-00083.safetensors 2.1 GBmodel-00073-of-00083.safetensors 2.1 GBmodel-00016-of-00083.safetensors 2.1 GBmodel-00008-of-00083.safetensors 2.1 GBmodel-00010-of-00083.safetensors 2.1 GB 2.1 GBmodel-00014-of-00083.safetensorsmodel-00033-of-00083.safetensors 2.1 GBmodel-00044-of-00083.safetensors 2.1 GBmodel-00068-of-00083.safetensors 2.1 GBmodel-00056-of-00083.safetensors 2.1 GBmodel-00034-of-00083.safetensors 2.1 GBmodel-00071-of-00083.safetensors 2.1 GBmodel-00027-of-00083.safetensors 2.1 GBmodel-00079-of-00083.safetensors 2.1 GBmodel-00042-of-00083.safetensors 2.1 GBmodel-00064-of-00083.safetensors 2.1 GBmodel-00049-of-00083.safetensors 2.1 GBmodel-00005-of-00083.safetensors 2.1 GBmodel-00012-of-00083.safetensors 2.1 GBmodel-00020-of-00083.safetensors 2.1 GBmodel-00080-of-00083.safetensors 2.1 GBmodel-00017-of-00083.safetensors 2.1 GBmodel-00030-of-00083.safetensors 2.1 GBmodel-00041-of-00083.safetensors 2.1 GBmodel-00054-of-00083.safetensors 2.1 GBmodel-00065-of-00083.safetensors 2.1 GBmodel-00078-of-00083.safetensors 2.1 GBmodel-00028-of-00083.safetensors 2.1 GBmodel-00052-of-00083.safetensors 2.1 GBmodel-00076-of-00083.safetensors 2.1 GBmodel-00063-of-00083.safetensors 2.1 GBmodel-00039-of-00083.safetensors 2.1 GBmodel-00050-of-00083.safetensors 2.1 GBmodel-00074-of-00083.safetensors 2.1 GBmodel-00026-of-00083.safetensors 2.1 GBmodel-00061-of-00083.safetensors 2.1 GBmodel-00037-of-00083.safetensors 2.1 GBmodel-00072-of-00083.safetensors 2.1 GBmodel-00024-of-00083.safetensors 2.1 GBmodel-00048-of-00083.safetensors 2.1 GBmodel-00059-of-00083.safetensors 2.1 GBmodel-00035-of-00083.safetensors 2.1 GBmodel-00022-of-00083.safetensors 2.1 GBmodel-00046-of-00083.safetensors 2.1 GBmodel-00070-of-00083.safetensors 2.1 GBmodel-00057-of-00083.safetensors 2.1 GBmodel-00004-of-00083.safetensors 2.1 GBmodel-00015-of-00083.safetensors 2.1 GBmodel-00013-of-00083.safetensors 2.1 GBmodel-00011-of-00083.safetensors 2.1 GBmodel-00009-of-00083.safetensors 2.1 GBmodel-00081-of-00083.safetensors 2.1 GBmodel-00007-of-00083.safetensors 2.1 GBmodel-00082-of-00083.safetensors 2.1 GBmodel-00003-of-00083.safetensors 2.1 GBmodel-00043-of-00083.safetensors 2.1 GBmodel-00006-of-00083.safetensors 2.1 GBmodel-00032-of-00083.safetensors 2.1 GBmodel-00067-of-00083.safetensors 2.1 GBmodel-00002-of-00083.safetensors 2.1 GBmodel-00019-of-00083.safetensors 2.0 GBmodel-00001-of-00083.safetensors 1.3 GBmodel-00083-of-00083.safetensors 842 MBinternlm/Sage-S2.1-Flash