whatmodel.sh / RWKV-4-Raven 14B
RWKV-4-Raven 14B
↗ share
+ contribute
RWKV-4-Raven 14B is an open-weights language model from RWKV (BlinkDL), released April 2023. Its published size is 14B parameters. Licence: apache-2.0.
Catalogue identity notes
Comparison-only. Arena entry 'RWKV-4-Raven-14B' (text). LMArena's leaderboard table (arena-leaderboard space @6f2eeeb) labels 'RWKV-4-Raven-14B' as 'RWKV-4-Raven-14B' (RWKV, Apache 2.0), linking https://huggingface.co/BlinkDL/rwkv-4-raven. Maker repo BlinkDL/rwkv-4-raven @b3ff339: licence tag apache-2.0. Parameters: nominal size from the model name. Released 2023-04: month the maker's Hugging Face repo was created (createdAt 2023-04-01). The repo holds several Raven versions/sizes; the Arena entry is the 14B model. RNN architecture, no single-command runner verified.
$ how it compares
The LLMs ranked either side of RWKV-4-Raven 14B, out of 828 with a score.
| # | model | score | where |
|---|---|---|---|
| 781 |  OpenLLaMA 3B v2 | 25±1 | local |
| 782 |  OPT-6.7B | 24±1 | local |
| 783 | RWKV-4-Raven 14B | 24±2 | comparison only |
| 784 |  Falcon3 1B Instruct | 24±4 | local |
| 785 |  StableLM-Base-Alpha 3B v2 | 24±1 | local |
$ what it's good at
chat & writing▁▃▁▃▅▄▆█▇▇▇▃ #453 of 463 · 98% of its score · 3 tests
Beats Phi-2 (2.7B), ChatGLM2-6B here, though they rank higher overall. Trails Gemma 3 270M IT, SmolLM2 360M Instruct here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Polyrating rating (bias-aware re-fit of Chatbot Arena votes, mid-2024) | 971.0 ✎ | ▂▁▂▃▃▇█▆▇▄▂▂ beats 12/95 | 34% |
| LMArena (text) | 1042.3 ✎ | ▁▃▁▂▃▃▃▆▆▇█▃ beats 7/379 | 34% |
| MT-Bench (GPT-4 single-answer grading, LMSYS runs) | 3.9800 MT-Bench score (1-10) ✎ | ▆▂▃▆▆▄██▂▄▄▆ beats 4/28 | 30% |
$ where to run it
No verified API or download recorded yet · add one
Weights and documentation: huggingface.co/BlinkDL ↗
$ how the score was worked out · Polyrating: A Cost-Effective… 34%, LMArena leaderboard dataset 34%, LMSYS Chatbot Arena leaderboa… 30%
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 Polyrating: A Cost-Effective and Bias-Aware Rating System for LLM Evaluation (arXiv:2409.00696v3) | 1 | 34% |
 LMArena leaderboard dataset | 1 | 34% |
 LMSYS Chatbot Arena leaderboard Space: MT-Bench column of the published leaderboard table | 1 | 30% |
| prior · pull toward the middle | 2% | |
| weighted average | = 24.2 |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json