User claims DeepSeek V4 Pro is mis-trained: cheaper Flash beats it
karminski3 · x · 2026-08-18
A Chinese user reports (unconfirmed) that in the DeepSeek V4 lineup, the much cheaper V4 Flash outperforms V4 Pro on reasoning, and the lower-effort V4 Pro High beats Pro Max. V4 Pro still writes better prose, but trails Opus 4.6 by a wide margin. The poster also asks for recommendations on writing benchmarks.
More from Models
- LiquidAI releases LFM2.5-VL-3B-WebGPU, a multimodal model optimized for browser deployment — LiquidAI · 2026-08-18
- McByte Sets New SOTA on SportsMOT Benchmark — NielsRogge · 2026-08-18
- Benchmark: Qwen 3.8 27B hits 50-60 t/s on dual RTX 5060 TI cards — chocofoxy · 2026-08-18
- Qwen 3.8 27B vs DeepSeek Flash: Benchmaxing or Real? — Best_Sail5 · 2026-08-18
- User reports Qwen model overthinking despite low reasoning effort setting — Jebbyk1 · 2026-08-18
- Qwen3.8-27B Uncensored MLX build trends on Hugging Face for Apple Silicon — orcarouter · 2026-08-18