FARE Models by Salesforce AI Research
Salesforce AI Research releases · 1 in the last 12 monthsChangelog →
1 model2025Up to 128K ctx
Last refreshed 2026-10-09. Next refresh: weekly.
Details
ResearcherSalesforce AI Research
LicenseCC-BY-NC-4.0
Commercial useNot permitted
Models1
Released2025
Max context128K
About
FARE (Foundational Automatic Reasoning Evaluators) is Salesforce AI Research's family of open-weight evaluator models, introduced in October 2025. They are trained with rejection-sampling fine-tuning on a large multi-task, multi-domain dataset to judge other models' outputs: pairwise comparisons, step-level error finding, reference-based and reference-free verification, and single-score ratings. FARE-20B is fine-tuned from OpenAI's gpt-oss-20b; the weights are released under CC BY-NC 4.0 for non-commercial use.
Decision facts
- Best fit
- General model comparison
- Capability starting point
- FARE-20B with 128K context
- Lowest tracked input
- Not tracked
- Closest related family
- SOV-Transcribe
Current Variants
Use-when guidance is based on each model's tracked capabilities, context window, release date, and replacement status.
1 in view
FARE-20BCurrent
Use when the workload needs 128K context and 20.9B parameters.
2025-10128K context20.9B parameters
| Model | Use when | Released | Signals | Status |
|---|---|---|---|---|
| FARE-20B | Use when the workload needs 128K context and 20.9B parameters. | 2025-10 | 128K context20.9B parameters | Current |
Release Timeline
1 release group2025-10
1 current
FARE-20B
Current128K context20.9B parameters
Specifications(1 models)
| Model | Released | Context | Parameters |
|---|---|---|---|
| FARE-20B | 2025-10 | 128K | 20.9B |
