model release

Shanghai AI Laboratory releases Sage S2.1 Flash, a 321B-parameter scientific multimodal model under MIT license

TL;DR

Shanghai Artificial Intelligence Laboratory has published Sage S2.1 Flash, a 321B-parameter scientific multimodal model with MIT-licensed weights and an FP8 variant. According to the lab, it placed first on four of 15 charted benchmarks and second on seven, including 90.3 on MLE-bench Lite.

3 min read
0

Shanghai Artificial Intelligence Laboratory has released Sage S2.1 Flash, a 321B-parameter scientific multimodal model, with code and weights under the MIT License. The model is listed on Hugging Face under the internlm organization, with an FP8 variant also available. Hugging Face and ModelScope repositories are provided.

What the model does

According to the model card, Sage S2.1 Flash is built for research tasks that require multiple rounds of analysis and verification. It combines scientific data understanding with information retrieval, programming and tool use. The lab says it can run specialized software, check results against intermediate outputs and visual feedback, and adjust later steps accordingly.

The Sage SciPerception architecture supports more than ten types of scientific input, including molecular graphs, crystal structures, and protein sequences and 3D structures. Scientific encoders and trainable projectors map each input type into a shared representation space, where it is processed alongside natural language instructions. Listed tasks include prediction, classification, site localization, quantitative analysis and structure inference.

Benchmark results (reported by the lab)

Across the 15 benchmarks and six models shown in the lab's chart, Sage S2.1 Flash placed first on four and second on seven. The comparison set is GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017, Qwen3.8-Max, GPT-5.6 Sol and Claude Opus 5.

First place:

  • MLE-bench Lite: 90.3 (DeepSeek-V4.1-Flash-1017: 89.2)
  • OfficeQA Pro: 72.2
  • Harvey LAB: 25.0
  • SpreadsheetBench 2: 37.1

Second place: AutomationBench, ALE-CLI, BFCL V4, WebArena-Verified, DeepResearch Bench II, JobBench and OSWorld-Verified.

Other scores from the results table:

Benchmark Sage S2.1 Flash Claude Opus 5 Qwen3.8-Max
Terminal-Bench 2.1 87.9 90.2 89.6
SWE-bench Pro 60.6 76.5 65.3
OSWorld-Verified 85.2 86.4 84.3
BFCL V4 75.5 80.6 66.8
BrowseComp 91.2 90.8 not reported

On agentic coding, Claude Opus 5 scores higher than Sage S2.1 Flash on all four listed benchmarks: Terminal-Bench 2.1, DeepSWE, NL2Repo-Bench and SWE-bench Pro. These are the lab's own figures.

Deployment

For general tasks such as coding, office productivity, security and computer use, the repository links separate deployment guides for the standard and FP8 versions. Scientific inputs require extra steps:

  • Download the projector from the model repository and the loading scripts from Sage-SciPerception.
  • Set up the scientific encoders, task-specific readouts and preprocessing dependencies separately.
  • Run the projector client-side, combine its output with the model's text token embeddings, and pass the result to the language model.
  • Enable --enable-prompt-embeds on the vLLM server and send prompt_embeds to /v1/completions, using the tokenizer and text embeddings that match the base model.

The weights are stored in Safetensors with F32 and BF16 tensor types. The page lists two quantizations. No Inference Provider currently deploys the model. Context window and pricing are not stated on the model page.

What this means

Sage S2.1 Flash pairs a general agentic toolset (coding, search, office and computer-use tasks) with scientific input handling through projectors. Its MIT-licensed weights and FP8 variant let teams self-host it. Using the scientific modalities takes more integration work than text-only use: separate encoders, readouts and preprocessing, plus prompt-embedding input over vLLM.

The benchmark claims come from the lab's own chart. The strongest results are in machine learning engineering and office and document tasks. On the agentic coding benchmarks in the table, the model trails Claude Opus 5, and on several it also trails Qwen3.8-Max and DeepSeek-V4.1-Flash-1017. Independent evaluation will be needed to confirm how these scores hold up outside the lab's setup.

Related Articles

model release

inclusionAI releases Duo 1.0 causal-base, a full-duplex multimodal world model initialized from MiniMax H3

inclusionAI has published the Duo 1.0 causal-base weights on Hugging Face. Duo is described as a full-duplex multimodal world model for real-time audio-visual interaction, initialized from MiniMax H3 and trained further by the Duo team.

model release

Meituan releases LongCat-Video-Avatar-1.6, open-weight avatar model for 10-minute-plus videos under MIT license

Meituan has released LongCat-Video-Avatar-1.6, an open-source audio-driven human video generation model with weights under the MIT License. The company claims it sustains videos longer than 10 minutes and ranks second among six systems in a human preference study of 33,051 valid votes.

model release

Qwen4 Max launches with 1M-token context and $1.50/$4.50 per 1M token pricing

Qwen4 Max, the flagship model in Alibaba's Qwen4 series, is listed on OpenRouter with a 1M-token context window. It accepts text, image and video input, returns text, and is priced at $1.50 per 1M input tokens and $4.50 per 1M output tokens.

model release

Meituan releases LongCat-2.5: 1.6T-parameter multimodal MoE with 48B active per token

Meituan's LongCat team has published LongCat-2.5, a native multimodal mixture-of-experts model with 1.6 trillion total parameters and approximately 48 billion activated per token. It targets long-horizon agentic tasks such as codebase development and full-project builds, and ships with W4A4 expert quantization for domestic AI accelerators.

Comments

Loading...