DeepSeek V4 Flash 0731 open weights released, outperforms flagship

DeepSeek-V4-Flash-0731, a 284B-parameter MoE model, scores 82.7 on Terminal-Bench 2.1, beating V4 Pro Preview's 72.1 despite fewer parameters. Priced near 2 cents per million input tokens, it jumped from 7% to 54% on DeepSweep agentic coding benchmark via post-training.
How this story unfolded
4 weeks · 4 reports · 15 community posts · from Aug 2
- Aug 2
- Aug 3
- Aug 4
- Aug 5
- Aug 20
- Aug 24
- Aug 25
- Aug 28
- Aug 30
AI Models by email
Get an email when there's news on AI Models
No news that day, no email.
More stories today
- Gradium TTS rises to #18 on Artificial Analysis Text-to-Speech Arena
- Tri Dao: More GPUs coming for open models
- Sentrux: Rust-based sensor helps AI agents improve code quality
- OpenClaw 2.0 launches with ClawHub, security concerns
- Google Antigravity introduces Boost deep reasoning