Skip to content
← Models
Open weightsShanghai AI Laboratory/10 trustVerified

InternVideo2.5

OpenGVLab's long-and-rich-context video MLLM: adaptive hierarchical compression plus preference optimization on dense visual annotations lets a 7B model memorize roughly 6x longer video inputs and track objects through them.

Visit official page →

About

InternVideo2.5 (arXiv 2501.12386, January 2025) is Shanghai AI Laboratory's upgrade of the InternVideo line, built on InternVL2.5 with what the authors call long and rich context (LRC) modeling. Two mechanisms carry it: adaptive hierarchical token compression (HiCo, the same idea as VideoChat-Flash from the same lab) for compact spatiotemporal representation, and direct preference optimization that transfers dense vision-task annotations into the MLLM. The payoff the authors report is a model that memorizes at least 6x longer video sequences than the base MLLM and picks up specialist vision skills like object tracking and segmentation. At 7B they report 75.7 on MVBench, 74.9 on PerceptionTest, 63.9 on EgoSchema, 60.6 on LongVideoBench, 72.8 on MLVU and 65.1 on VideoMME — strong numbers, all author-reported. Code and models are Apache 2.0 at OpenGVLab/InternVideo (InternVideo2.5 subdirectory) with Hugging Face checkpoints. No hosted API.

Specs

Context window

At least 6x longer input sequences than the base InternVL2.5 (author-reported); 5,000-frame needle-in-a-haystack eval demonstrated.

Pricing

Free / open weights — Apache 2.0

Free quota

Fully open — code and checkpoints on GitHub / Hugging Face; self-host, no registration

API access

No hosted API — self-host via GitHub / Hugging Face checkpoints

Weights

Open weights

Released

Not disclosed

Benchmarks

  • 7B: MVBench 75.7, PerceptionTest 74.9, EgoSchema 63.9, LongVideoBench 60.6, MLVU 72.8, VideoMME 65.1 (author-reported)\n- Reported gains of +3.7 to +12.4 points over the InternVL2.5 base across benchmarks (author-reported)\n- No independent replication of the headline figures found

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

long-video-qavideo-qaobject-trackingvideo-summarizationvideo-retrievalvideo-segmentation

Trust score — 8/10

/10 trustVerified

Scored 8/10 (TIER verified, metrics-v1). Evidence: technical report with Apache 2.0 code and public checkpoints; 2,400+ stars on the InternVideo repo; the LRC design is concrete with compatibility ablations (HiCo plus preference optimization); from the lab behind InternVL and VideoChat-Flash. Strengths: strong reported results at 7B, genuinely longer effective context, bonus specialist skills like tracking. Docked for: (1) all benchmark figures are author-reported with no independent replication; (2) it is an incremental upgrade on InternVL2.5 rather than a from-scratch architecture. License: Apache 2.0.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
8.0/10
Maintenance · recency, releases, issue load
9.0/10
Openness · license + docs
10.0/10
Manual review · sanity read + red-flag check
8.0/10
  • · OpenGVLab long-and-rich-context upgrade of InternVL2.5: HiCo compression + DPO on dense annotations.
  • · 7B: MVBench 75.7, MLVU 72.8, VideoMME 65.1 (author-reported); ~6x longer inputs.
  • · Bonus specialist skills: object tracking and segmentation.

Sources