Ray 2.0
Ray 2.0 is Wordfully's AI writing model for grammar correction and rephrasing. The evaluation below looks at the quality of its edits and how quickly it returns a suggestion.
- 48 ms to first correction
Reported in Wordfully's internal evaluation. Latency measured on a 50-word passage.
Grammar correction with context.
Ray 2.0 is a specialized text-editing model. Wordfully's technical report describes training on editorial text and human revision examples, with two priorities: correcting language errors and preserving meaning when wording changes.
Our instant correction model currently supports English only. The Wordfully team is working to add more languages. Rephrasing, translation, and other writing tools already work in most widely spoken languages.
Grammar correction needs context. Changing a verb can shift when something happened, and replacing "may" with "will" can make a statement more certain than the writer intended. A useful edit needs to preserve those details.
Rephrasing gives a sentence a different structure or tone. It can make a dense paragraph easier to follow, but the names, facts, and point of the original still matter. Ray's evaluation covers correction and rewriting as separate tasks so you can see how it performs in each.
Ahead of every model in our comparison.
In Wordfully's tests, Ray 2.0 led the compared open models on grammar correction, rephrasing, and time to the first correction. The tables below show the reported results.
Grammar correction
| Benchmark | Ray 2.0 | CoEdIT-XXL | CoEdIT-XL | GECToR-RoBERTa | Flan-T5 Large | T5 Base | Llama 3.3 70B |
|---|---|---|---|---|---|---|---|
| BEA-2019 · F₀.₅ | 81.6 | 77.2 | 74.8 | 72.3 | 68.4 | 62.1 | 75.5 |
| CoNLL-2014 · F₀.₅ | 73.4 | 69.1 | 66.8 | 65.4 | 61.2 | 54.7 | 67.3 |
| False-positive edits ↓ | 1.2% | 3.4% | 4.6% | 4.8% | 7.1% | 11.4% | 8.4% |
Scroll to compare all models
BEA-2019 uses ERRANT; CoNLL-2014 uses M². False positives are edits to clean sentences in the FCE test split. Flan-T5 Large is pszemraj's grammar-synthesis model; T5 Base is Vennify's grammar-correction model. GECToR-RoBERTa's exact checkpoint is unverified.
Rephrasing and fluency
| Benchmark | Ray 2.0 | DIPPER-XXL | CoEdIT-XXL | Palmyra 20B | Humarin T5 | Parrot T5 | Qwen 2.5 72B |
|---|---|---|---|---|---|---|---|
| JFLEG · GLEU | 68.2 | 63.8 | 64.6 | 63.1 | 59.4 | 58.1 | 62.5 |
| EditEval · SARI | 54.9 | 51.8 | 51.5 | 49.6 | 47.2 | 46.1 | 50.1 |
Scroll to compare all models
These results compare model checkpoints, including Grammarly's CoEdIT. They do not compare the commercial Grammarly, QuillBot, or Writer applications.
Speed that keeps up with a sentence.
| Response time | Ray 2.0 | T5 Base | Humarin T5 | CoEdIT-XL | CoEdIT-XXL | DIPPER-XXL | Palmyra 20B | Llama 3.3 70B |
|---|---|---|---|---|---|---|---|---|
| First correction | 48 ms | 65 ms | 78 ms | 135 ms | 240 ms | 260 ms | 310 ms | 450 ms |
| Full response | 112 ms | 190 ms | 215 ms | 380 ms | 680 ms | 720 ms | 890 ms | 1,520 ms |
Scroll to compare all models
First-correction time measures when a suggestion starts to arrive, not when the full response is ready. Hardware, batch size, and latency percentiles are not specified in the source report.
About these results
Internal source: Wordfully Ray 2.0 Technical Benchmark & Evaluation Report, WFD-EVAL-2026-RAY2-01.
The report does not specify test splits or sample counts for BEA-2019, CoNLL-2014, JFLEG, or EditEval. It also omits model revision IDs, decoding settings, and raw run outputs. It mentions partner access to checkpoints and synthetic evaluation splits. Results describe only the listed models and reported conditions, not every model on Hugging Face.
Selected model documentation: CoEdIT-XXL, DIPPER-XXL, Vennify T5
An AI writing assistant for everyday work.
Wordfully brings grammar checking and rewriting tools into supported writing fields, including Google Docs, Gmail, and Slack. You can review a suggestion where it applies and keep the wording that fits your message.
Grammar checking is useful when you want to correct a sentence without changing its purpose. Rephrasing helps when the wording needs more work: a shorter email, a clearer explanation, or a different tone for the reader.
The extension's model picker currently offers models such as GPT, Claude, and Gemini. Ray's benchmark scores refer to the model evaluated on this page, not to those rephrasing options.
The final edit is yours. Read the suggested version in its original context, check that it keeps your meaning, and choose the wording you want to use.
See Wordfully's AI writing tools