SSyncDev

Intern-S2-Mobius

SyncDev model directory · Profile updated Nov 18, 2026

Compare with another model

Intern-S2-Mobius is Shanghai AI Lab's 35B model that separates a shared knowledge memory from reasoning modules, reporting about 4x faster inference.

Key specifications

Specification score
44.6
Context window
262.14K
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: No API pricing stated; open weights for self-hosting

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input
  • Image input

Produces

  • Text output

Capabilities

ReasoningVisionOpen weights

What is Intern-S2-Mobius?

Every Transformer layer does two jobs at once: it stores facts in its feed-forward weights and it reasons over the hidden state. Intern-S2-Mobius pulls those jobs apart. Knowledge lives in one globally shared Memory, and several Reasoners repeatedly query and refine the hidden state against it.

The result, released by Shanghai AI Laboratory on 29 July 2026, is a 35B foundation model on a new Mobius-v0 architecture, implemented with XTuner and LMDeploy. It was continually pre-trained from Qwen3.5-35B, then supervised-tuned and reinforced.

What the design buys

  • Backward Residual Connection. Early and late reasoning stages can reach knowledge outside their own layer, rather than only what flowed forward.
  • Dynamic Latent Reasoning. The model iterates on continuous hidden states before it decodes. Some of the deliberation that would normally appear as visible chain-of-thought happens internally, and compute is spread unevenly across tokens.
  • Shorter traces. Reasoning outputs are markedly shorter than the Qwen3.5-35B baseline's, which is where most of the reported speedup comes from.

Shanghai AI Lab reports comparable or better scores than Qwen3.5 35B-A3B on general reasoning and close to 4x end-to-end inference speedup, with large gains on Biology-Instructions, Mol-Instructions and MolecularIQ. Those are the publisher's measurements.

Numbers for planners

Field Value
Hub repo internlm/Intern-S2-Mobius (FP8 build published 4 August)
Parameters About 36B in safetensors
Memory experts 2,560, with 8 selected per step
Layers 40
Positions 262,144
Modalities Text and images in, text out
Licence Apache-2.0

Evaluation used 64K inference length on MMLU-Pro, SimpleQA and HLE and 128K on the other text benchmarks.

A caution on tooling

A new architecture means new kernels. Check that your serving stack supports the interns2_mobius model type before committing; LMDeploy is the reference path. Generic Transformer optimisations may not apply.

Where it fits

Choose it when throughput per GPU matters and you can test an unusual architecture. Measure tokens per answer and requests per second on your own prompts next to the Qwen3.5 baseline, and check that shorter reasoning does not cost accuracy on your hardest cases. For peak capability rather than efficiency, the lab's larger Intern-S2-397B is the comparison point; both sit under Shanghai AI Lab's Hugging Face listings.

How to evaluate Intern-S2-Mobius

Compare the model on the work you need to complete.

Tasks and inputs

Intern-S2-Mobius is listed for Image-text-to-text, Text generation. It accepts text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Documented limits: 262.14K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

It replaces layer-bound feed-forward knowledge with a globally shared Memory queried by multiple Reasoners, adding backward residual connections and latent reasoning.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

sensetime

SenseNova-Loop-9B

SenseNova-Loop-9B is SenseTime's 9B looped multimodal model: 32 layers run as 56, 512K context, ~15 GiB at 256K in 4-bit, open weights under Apache 2.0.

524.29K context

Tencent

Hy4 Flash

Tencent Hy4 Flash is the low-cost Hy4 tier with open weights: a 220B MoE (~11.3B active) for agents and documents, with 1M context and image input.

1.05M context

shanghai-ai-lab

Sage S2.1 Flash

Sage S2.1 Flash is Shanghai AI Lab's open scientific multimodal agent model, reading molecules, crystals and proteins alongside text and images.

— context

Meituan

LongCat-2.5

LongCat-2.5 is Meituan's open-weight MIT-licensed multimodal MoE for long-horizon agents: screenshots, charts, video and multi-step coding.

— context

sensetime

SenseNova-Loop-9B-Preview

SenseNova-Loop-9B-Preview is SenseTime's looped 9B text-and-image model for long-context agentic work, running shared layers twice for extra depth.

262.14K context

Liquid AI

LFM3-VL-2B

LFM3-VL-2B is Liquid AI's open 2B vision-language model for images and video on-device, strong on documents, charts and GUI screenshots.

— context