Original Title: internlm/Sage-S2.1-Flash
A scientific multimodal model uploaded by the Shanghai Artificial Intelligence Laboratory on 11/15. The deployment documentation states it is a Mixture of Experts (MoE) model with 320B total parameters and 18B active parameters, utilizing the GLM-5.3-Flash architecture. The BF16 weights are approximately 645 GB. It integrates scientific data understanding with information retrieval, programming, and tool calling, enabling it to invoke specialized software for analysis and adjust subsequent steps based on intermediate results and visual feedback. Through the Sage SciPerception projector, it can ingest over ten types of scientific inputs, including molecular graphs, crystal structures, protein sequences, and 3D structures. An FP8 version is also available, with code and weights licensed under MIT.
Why It Matters Integrating the understanding and retrieval of over ten types of scientific data, along with programming and software invocation, into a single model makes it suitable for research workflows that require repeated analysis and verification.
Rising Score 55 · Likes 55 · Downloads 22
Reasons
- Officially claimed to achieve first place in 4 items and second place in 7 items across the 15 benchmarks listed in the chart, among 6 models: MLE-bench Lite 90.3 (DeepSeek-V4.1-Flash-1017 scored 89.2), OfficeQA Pro 72.2, Harvey LAB 25.0, and SpreadsheetBench 2 scored 37.1.
- Scientific data is processed via the projector: inputs such as molecules, crystals, and proteins are first encoded and then mapped into the language model's input space, enabling property prediction, classification, site localization, quantitative analysis, and structural inference.
- MIT licensed, available for download on Hugging Face and ModelScope. An FP8 version is provided, along with deployment instructions for LMDeploy, vLLM, and SGLang.
Usage
- General tasks (programming, office work, computer operation): Install vLLM 0.33.0, SGLang 0.5.20+, or LMDeploy following the official GLM-5.3-Flash recipe. Replace the model name with internlm/Sage-S2.1-Flash and add --trust-remote-code for each engine; the official example uses 8 H200 GPUs (--tensor-parallel-size 8).
- For vLLM, first run hf download internlm/Sage-S2.1-Flash --include "vllm_patch/*" to retrieve the patch and apply it according to its README (fixing issues where images are mismatched with results during parallel tool calls). For SGLang, use --chat-template to specify the chat_template.jinja included in the repository.
- Scientific inputs: Download projectors/sage-s2.1-flash/ from this repository, and download load_projector.py and modeling_projector.py from internlm/Sage-SciPerception. Use the projector to generate prompt_embeds, then send them to /v1/completions with --enable-prompt-embeds enabled in vLLM.
Risks
- Large size: BF16 weights are approximately 645 GB, and the official example uses 8 H200 GPUs, making it unrunnable on personal computers; FP8 KV cache is not supported on H100 / H200.
- All scores are derived from official charts; performance in several programming benchmarks is lower than comparison models. For example, SWE-bench Pro 60.6 and NL2Repo-Bench 57.2 both score lower than Claude Opus 5 and DeepSeek-V4.1-Flash-1017.
- Omitting --trust-remote-code does not trigger an error but causes strings like <tool_call> to be treated as control tokens, leading to premature stopping or tool call errors; there are also known issues with non-streaming responses in vLLM.
- Scientific inputs require installing separate scientific encoders and preprocessing dependencies, making the workflow more complex than that of standard conversational models; this entry has not been executed.