InternLM

Shanghai AI Lab (上海人工智能实验室) has published InternLM weights since 2023, but the line that matters now is Intern-S, aimed at scientific reasoning rather than general chat. Intern-S1-Pro, open-sourced in February 2026, runs 1 trillion total parameters across 512 experts with 8 activated per token for roughly 22B active. Intern-S2 followed through mid-2026 with a 397B preview checkpoint and a 35B Mobius variant continually pretrained from Qwen3.5-35B. All of it ships Apache 2.0.

The architectural choices are unusual enough to be worth reading the technical report for. The vision path is trained on raw pages of scientific literature so formulas, figures and surrounding text are modelled jointly rather than parsed into plain text first, which preserves the spatial correspondence that gets destroyed by a PDF-to-text pipeline. Fourier Position Encoding and a rebuilt sequence encoder target long physical signals, quoted as supported to roughly 10^6 points. On SciReasoner the lab reports 55.5 for Intern-S1-Pro. Intern-S2-Mobius reports scores at or above its Qwen3.5-35B baseline on general reasoning while claiming close to 4x end-to-end inference speedup from shorter reasoning traces.

The practical questions resolve well. Weights sit on both Hugging Face and ModelScope, FP8 checkpoints are published alongside BF16, and LMDeploy, vLLM, SGLang and Transformers all have documented paths with MTP speculative decoding recommended for Mobius. The lab's statement that Intern-S1-Pro was trained on domestic computing infrastructure is a procurement signal worth weighing if you are planning for Ascend hardware, though it is a statement about training rather than a guarantee of inference support. The frustrations here are organisational rather than technical: too many concurrent model names, Preview tags on the flagships and English docs that arrive after the Chinese.