SlowFast-LLaVA
Apple's training-free video LLM: a two-stream SlowFast design feeds a few high-resolution frames and many low-resolution frames into an existing LLaVA, capturing detail and motion without any video training at all — plus a trained 1.5 family with open weights.
About
SlowFast-LLaVA (arXiv 2407.15841, July 2024) from Apple starts from a contrarian premise: you can get a strong video LLM without training on video. The Slow pathway samples frames at a low rate but keeps full spatial detail, while the Fast pathway samples at a high rate with heavy spatial pooling to capture motion cues; together they fit both into a normal LLM token budget. The training-free version (SF-LLaVA) plugs into existing LLaVA-NeXT weights and still beat prior training-free methods on every tested benchmark, with the 34B variant even edging video-fine-tuned models on EgoSchema (author-reported). The follow-up SlowFast-LLaVA-1.5 (arXiv 2503.18943, March 2025) adds a real trained family at 1B, 3B and 7B with a deliberately simple two-stage recipe, reporting 62.5% on LongVideoBench and 71.5% on MLVU at 7B — and Apple released all weights and training datasets. The v1 repo (apple/ml-slowfast-llava) has no license file detected, so verify terms; the 1.5 weights are public. No hosted API.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- SF-LLaVA-34B (training-free): beat the prior training-free SOTA on all tested benchmarks; +11.4% over IG-VLM on EgoSchema at 7B (author-reported)\n- SF-LLaVA-1.5-7B: 62.5% LongVideoBench, 71.5% MLVU, 60.8% Video-MME (w/o subtitles) at 3B (author-reported)\n- All figures author-reported; no independent replication found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (TIER verified, metrics-v1). Evidence: Apple-published papers with public code; the 1.5 family ships open weights and open training datasets; the SlowFast two-stream idea is elegant, well-ablated and widely referenced. Strengths: a genuinely useful training-free baseline plus a properly trained efficient family, with unusual transparency on data. Docked for: (1) no license file detected in the v1 repo, so terms need checking; (2) all benchmark figures are author-reported. License: none detected in v1 repo — verify before commercial use.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Apple's training-free two-stream (slow hi-res + fast lo-res) video LLM design.
- · SF-LLaVA-1.5: 1B/3B/7B trained family with open weights and open training datasets.
- · LongVideoBench 62.5%, MLVU 71.5% at 7B (author-reported).