DeepSeek V4 Models Show Similar Evaluation Results
Teortaxes shares observations on performance consistency among updated versions.
Teortaxes, known for promoting DeepSeek, shared a detailed comparison of the company's V4 models. According to the post, these updated models reach nearly identical standings on various third party evaluations. The commentary highlights that results align closely with those from Flash-0731 while suggesting similar computational resources are involved. The author questions the value provided by substantially larger parameter counts in the latest iterations, noting a lack of clear gains from the added scale.
Combined views
3.1K
1 post, first seen 1h ago
DeepSeek V4 Models Show Similar Evaluation Results
Teortaxes shares observations on performance consistency among updated versions.
Teortaxes, known for promoting DeepSeek, shared a detailed comparison of the company's V4 models. According to the post, these updated models reach nearly identical standings on various third party evaluations. The commentary highlights that results align closely with those from Flash-0731 while suggesting similar computational resources are involved. The author questions the value provided by substantially larger parameter counts in the latest iterations, noting a lack of clear gains from the added scale.
Reactions
1 postT( 
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTexinexplicable stuff it's not just "comparable", it's precisely 0.1% lower than Flash-0731, and the cost difference suggests comparable token budgets. All updated V4 models are ≈same level on most third party evals. At present not clear what 5.6x more total/3x active params buy.
- likes: 22
- replies: 4
- bookmarks: 4
- reposts: 0