InternLM 2.5

weights

open(custom 'Free Commercial License', academic free, commercial use free but requires an application form)

data

closed(pretraining corpus not released)

code

open(Apache-2.0 training/inference code on GitHub)

license

Apache-2.0(code, OSI) + custom weights license(non-OSI, application step)

The training and inference code is Apache-2.0, but the weights ship under a custom, non-OSI 'Free Commercial License': free for academic work and free for commercial use, but only once an application form has been approved. That application step on the weights, together with a pretraining corpus that is not released, is what keeps this at open weights rather than open source.

• https://api.github.com/repos/InternLM/InternLM recorded 2026-08-13

  `private: false`, `archived: false`, license `Apache-2.0`, 7,264 stars, description "Official release of InternLM series (InternLM, InternLM2, InternLM2.5, InternLM3)".

• https://huggingface.co/api/models/internlm/internlm2_5-7b recorded 2026-09-29

  Hub API record for internlm/internlm2_5-7b with gated false and private false, weight shards "model-00001-of-00002.safetensors" and "model-00002-of-00002.safetensors" listed. The record carries no license tag.

• https://huggingface.co/internlm/internlm2_5-7b recorded 2026-09-29

  Model card license section - "The code is licensed under Apache-2.0, while model weights are fully open for academic research and also allow" free commercial usage, and "To apply for a commercial license, please fill in the" application form. No pretraining corpus is named or linked.

• https://huggingface.co/internlm/internlm2_5-7b/raw/main/README.md recorded 2026-09-29

  Raw model card - "The code is licensed under Apache-2.0, while model weights are fully open for academic research and also allow" free commercial use on application, and the corpus is described only as "InternLM2.5 leverages a large amount of synthetic data". Searched the card for dataset: none.

• https://pypi.org/pypi/InternEvo/json recorded 2026-09-29

  PyPI record for InternEvo, InternLM's training framework, summarized as "an open-sourced lightweight training framework aims to support model pre-training without the need for extensive dependencies". The record declares no license field.

Adoption is measured as Hugging Face downloads for the single declared checkpoint, internlm2_5-7b.

• https://huggingface.co/api/models/internlm/internlm2_5-7b recorded 2026-08-13

  3,095 downloads in the trailing 30 days for internlm/internlm2_5-7b

InternLM2.5-7B scores 71.6 on MMLU (5-shot, OpenCompass), matching Yi-1.5-9B and beating Llama-3-8B's 66.4, with a 1M-token context. That places it beside Falcon 3 and Yi-1.5 among small models from the same generation, well below the 2026 frontier.

• https://huggingface.co/internlm/internlm2_5-7b recorded 2026-09-29

  Model card benchmark section - "We have evaluated InternLM2.5 on several important benchmarks using the open-source evaluation tool" OpenCompass; the table row reports MMLU 71.6 for InternLM2.5-7B against 65.8 for InternLM2-7B, 66.4 for LLaMA3-8B and 71.6 for Yi-1.5-9B.

• https://huggingface.co/internlm/internlm2_5-7b/raw/main/README.md recorded 2026-09-29

  Raw model card table - "| MMLU | 71.6 | 65.8 | 66.4 | 71.6 |" under the header "| Benchmark | InternLM2.5-7B | InternLM2-7B | LLaMA3-8B | Yi-1.5-9B |", results "obtained from" OpenCompass.

• https://raw.githubusercontent.com/InternLM/InternLM/main/model_cards/internlm2.5_7b.md recorded 2026-08-13

  InternLM2.5-7B base model card, MMLU 71.6 (OpenCompass). Cited previously as a github.com blob URL; the raw form is what the fetcher reads and what the digest is taken over.