PerceptionLM
Fully open, reproducible vision-language models (1B/3B/8B) for image and video understanding, with no proprietary distillation.
About
PerceptionLM is Meta FAIR's attempt at a fully transparent vision-language model family: 1B, 3B and 8B models trained with no distillation from proprietary models, with the data, training recipes and evaluation code all released. The paper (April 2025) centers on detailed video understanding, filling what the authors identify as data gaps in spatio-temporal reasoning with 2.8M newly collected human-labeled video question-answer and caption instances, plus a new PLM-VideoBench evaluation suite.
The headline claim is that a sub-8B model on fully open data can compete with systems like InternVL3 and Qwen2.5-VL. That is author-reported and worth independent checking, but the release itself is unusually complete for reproducibility research. The weights sit behind a gated Hugging Face page under the FAIR research license, so this is a research artifact, not a commercial base model. For labs that care about provenance and replicability in video understanding, it is one of the most serious open efforts available.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Author-reported competitive results vs InternVL3 / Qwen2.5-VL (paper, arXiv 2504.13180); no independent eval found
- PLM-VideoBench introduced by the authors for what/where/when/how video reasoning
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER metrics, metrics-v1). Evidence: arXiv 2504.13180 (Apr 2025) from Meta FAIR; HF facebook/Perception-LM-1B (1B/3B/8B family); github.com/facebookresearch/perception_models with training, eval and data; 2.8M human-labeled video instances released. Strengths: fully open data/code/weights pipeline, no proprietary distillation, new PLM-VideoBench suite. Docked for: (1) FAIR research license, not commercial-open, plus gated access; (2) competitive claims are author-reported only. License: FAIR Research License.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · arXiv 2504.13180 (2025-04-17) from Meta FAIR; 1B/3B/8B family
- · github.com/facebookresearch/perception_models: training, eval, data, PLM-VideoBench all released
- · 2.8M human-labeled video QA/caption instances; no proprietary-model distillation (paper)
- · Weights gated behind HF terms under FAIR Research License (research only, not commercial-open)
- · Competitive claims vs InternVL3/Qwen2.5-VL are author-reported only