Estimating True Parameter Sizes of DeepSeek V4 and Kimi K3
teortaxesTex · x · 2026-08-05
A technical blogger used a heuristic formula (square root of activetotal parameters) to estimate the Dense Equivalent Power Level of frontier models: V4-Flash is 60B, GLM is 172B, V4-Pro is 280B, and Kimi K3 is 540B.
The author notes that under the old regime of independent RL projects, the updated V4-Pro would likely outperform K3. However, since much of V4-Pro's uplift was crammed into the Flash variant, its actual ceiling needs to be readjusted.
More from Models
- Alibaba's Qwen3.8-Max Leaks: 2.4T Parameters and 1M Context — Aiden_Tech_Ai · 2026-08-05
- Developer accidentally tweaks Gemma layer, changes its favorite color — dejanseo · 2026-08-05
- Developer: DeepSeek Flash is genuinely good, failures are just intellectual limits — yacineMTB · 2026-08-05
- Alt sources for MiniMax H3 quants and LoRAs — reeight · 2026-08-05
- Ridiculous Experience: User Kicked from Fable to Sonnet to Haiku — basedjensen · 2026-08-05
- Xiaomi Hires DeepSeek Core Researcher Luo Fuli with Tens of Millions Salary to Lead AI — thisdudelikesAI · 2026-08-05