CoS-VQ BF16 - WikiText-2 PPL: leaderboard
The CoS-VQ study reports these reference results for unquantized models on language modeling and zero-shot questions. Lower perplexity is better. Perplexity (sequence length 2048).
Metric: Perplexity (sequence length 2048). Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.1 70B Instruct | 3.78 | #559 |
| 2 | Llama 3.1 8B Instruct | 7.22 | #1035 |
| 3 | Qwen 3 32B | 7.61 | #443 |
| 4 | Qwen 3 14B | 8.65 | #536 |
| 5 | Qwen 3 8B | 9.72 | #671 |
| 6 | Qwen 3 4B | 13.66 | #833 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=cos-vq-bf16-wikitext-2-ppl · How It Works · Data refreshed daily, snapshot 2026-10-29.