---

license: apache-2.0

base_model:

- internlm/Intern-S2-397B

base_model_relation: quantized

pipeline_tag: text-generation

library_name: gguf

language:

- zh

- en

tags:

- science

- moe

- gguf

- llama.cpp

- 42model

---

<!-- README for the Hugging Face repo 42ailab/Intern-S2-397B-GGUF. Editing this file changes the repo landing page. -->

Intern-S2-397B · A Science Model That Runs on One 128 GB Machine

English | 简体中文 · Also on ModelScope · 42ailab.com

A science model of nearly 400 billion parameters, compressed to about 98 GB so that it runs offline on a single computer with 128 GB of memory: read papers, check derivations, write research code, run long research tasks. Nothing leaves your machine.

> The model was built by Shanghai AI Laboratory (InternLM team) and released under Apache-2.0 (Intern-S2-397B). This repository is not a new model: it is a quantized GGUF we made from the official weights, plus our own test results. The capability belongs upstream.

What it is for

Upstream describes Intern-S2-397B as its most capable model for scientific intelligence and long-horizon agents, trained with large-scale reinforcement learning across more than 20 scientific domains. The official weights are about 800 GB in full precision (about 400 GB in the official 8-bit release), which normally means a multi-GPU server.

This build is about 98 GB, so a Mac or PC with 128 GB of memory can run it locally.

How we compressed it

We started from the official 8-bit release and brought it down to about 1.95 bits per weight on average. What matters is where the bits go, not how hard everything is squeezed:

• The model is a mixture of experts: 512 expert blocks, of which 10 are used per token. The experts are about 90% of the size, so only they get extreme compression.
• Attention and the other small but quality-sensitive parts stay at 8-bit.

We also tried the common approach of pushing every part to about 1 bit: the model fell into repetition and nonsense. With the allocation above, the same four test prompts were answered normally.

What we measured

For the model's benchmark results, see the upstream model card. The numbers below are our own checks on an Apple M3 Max with 128 GB:

| Check | Result |

|---|---|

| On Chinese technical text the model had not seen, how often its top prediction matches the 8-bit version | 81% of positions |

| Perplexity on the same text, relative to the 8-bit version (lower is better) | about 9% higher |

| Four prompts (Chinese explanation, Chinese writing, a time calculation, Python) | all answered normally; the calculation and the code were correct |

| Generation speed (GPU) | about 6–15 tokens per second |

These are small samples (one run per prompt), not official benchmark scores.

Limitations

• Needs 128 GB of memory. The files are about 98 GB, and running needs more on top.
• Text only. The upstream model also reads images; this build does not include the vision part.
• Thinks at length first. Upstream enables thinking by default, so answers are preceded by a long reasoning trace; you can turn thinking off for simple questions.
• Extreme compression costs something. About 1 in 5 top predictions differ from the 8-bit version. Where precision matters most, use the official weights.

Use

Packaged for 42model: in the desktop app, open Models → Research and download Intern-S2-397B; or run 42model download intern-s2:397b-iq1_s.

Files and license

| File | Role | Size |

|---|---|---:|

| Intern-S2-397B-IQ1_S-mix-00001-of-00003.gguf and 2 more | the model, split to fit the per-file size limit (loaded together automatically) | 91.4 GiB total |

Weights are under the upstream Apache-2.0 license (official weights). Upstream does not give a citation for this model; its predecessor's technical report is Intern-S2-Preview (arXiv:2608.13505), which describes the Preview release, not this one.