ReasonIR-8B
Meta's 8B bi-encoder built for reasoning-heavy retrieval, SOTA on BRIGHT at release.
About
ReasonIR-8B is the first retriever explicitly trained for reasoning tasks rather than short factual lookup. Released by Meta in April 2025 (paper at COLM 2025), it fine-tunes LLaMA-3.1-8B on a mix of public data and synthetic data from the ReasonIR-Synthesizer pipeline, which generates challenging queries and hard negatives that force the model to reason rather than pattern-match.
On the BRIGHT reasoning-intensive retrieval benchmark the authors report 29.9 nDCG@10 standalone and 36.9 with a lightweight reranker, beating far more expensive reranker baselines at roughly 200x lower inference compute. In RAG setups it lifted MMLU by 6.4% and GPQA by 22.6% over the closed-book baseline. The catch is the CC-BY-NC-4.0 license: the weights are free for research but commercial use is barred.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- BRIGHT: 29.9 nDCG@10 standalone, 36.9 with Qwen2.5 reranker (author-reported, SOTA at release)
- RAG: +6.4% MMLU, +22.6% GPQA vs closed-book baseline (author-reported)
- Beats Rank1-32B-class rerankers at ~200x lower inference compute (author-reported)
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 (verified tier, metrics-v1). Evidence: HF API confirms CC-BY-NC-4.0 license and April 2025 release; facebookresearch/ReasonIR repo and open training data published; COLM 2025 paper. Strengths: Meta research pedigree, open code/data/model, genuinely new capability (reasoning retrieval). Docked for: (1) NC license excludes all commercial use; (2) every benchmark figure is author-run; (3) only ~669 HF downloads, so adoption is research-only.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 4.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · HF API confirms CC-BY-NC-4.0 and April 2025 release; facebookresearch/ReasonIR repo + open training data
- · COLM 2025 paper; BRIGHT 29.9 standalone / 36.9 with reranker are author-reported SOTA claims
- · Only ~669 HF downloads; adoption is research-only