Deepseek V4 Flash hits 420 tok/s in new community benchmark
HankYeomans · x · 2026-08-27
A community member achieved a new throughput record for Deepseek V4 Flash through specific configuration tuning.
- Performance Milestone: Single-stream decode speed reached 420 tokens/s, surpassing the previous community record of 357.9 tok/s.
- Optimization Recipe: By adjusting two variables and applying a 6000MHz memory overclock, the user peaked at 446.4 tok/s.
- Config Details: The run utilized TP=4 (Tensor Parallelism=4) with prefill disabled to focus purely on decoding performance.
More from Models
- Test: Codex Usage Seems Almost Unlimited Compared to Fable — panickssery · 2026-08-27
- Qwen and GLM push forward with low-cost, high-speed models — brandon_galang · 2026-08-27
- User Corrects Misinformation Regarding Qwen3.8-27B Model Rankings — Real-C- · 2026-08-27
- Rumor Suggests Imminent Release of Anthropic's Sonnet and Opus — ChrisGPT · 2026-08-27
- Qwen3.8-27B on AMD R9700 hits 227 tok/s with lossless block-diffusion drafter — samsja19 · 2026-08-27
- Rumor: Anthropic to launch Fable 5.1 before Astra is ready — haider1 · 2026-08-27