Key Info

Requesty ran Particle-1, a post-trained version of DeepSeek V4.1 Flash, against its base model on an internal coding benchmark of 88 real Go backend tasks with hidden tests and the same harness.

Highlights

  • 39% fewer tokens per task, 20% lower cost, and half the time per turn versus the base model
  • 2.7s median per turn, the fastest of 19 models tested, ahead of glm-5.3-flash (3.4s) and gpt-6-sol (3.6s)
  • Trade-off: 61% of tasks solved vs 67% for its base, level with DeepSeek V4 Pro among open-weight models
  • Frontier models still lead; Opus 5.5 solved 90%
  • Positioned for high-volume coding loops where token and turn-time costs accumulate; available as model="particle/particle-1"