DeepSeek V4 Flash Weights Released: Runnable on Single GPU, Codex Support
tokenbender · x · 2026-07-31
DeepSeek V4 Flash 0731 weights are now available on Hugging Face. This version is a pure post-train boost over the April checkpoint, features agent-native executor, Codex support, and includes dspark specdec module. Strongly quantized weights may run on a single GPU, and it is reported efficient and intelligent per AA report.
Related event: DeepSeek-V4-Flash Enters Public Beta with Enhanced Agent Capabilities(73 posts)→
More from Models
- DeepMind's GROD2 Adapts to New Robot Bodies in Under Two Hours — DynamicWebPaige · 2026-08-01
- Claude 3.5 Helps Disprove 80-Year-Old Jacobian Conjecture — gregd_nlp · 2026-08-01
- ARC-AGI-3 Replay: Claude Opus Beats GPT-5.6 via Durable State Tracking — otarU · 2026-08-01
- Red Hat Releases Quantized Kimi K3: FP4/FP8 for Accelerated Inference with Minimal Quality Loss — _akhaliq · 2026-08-01
- Google AI Recap: Gemini Robotics 2, New Flash Models, and Music Generation — GoogleAI · 2026-08-01
- Researcher Recaps Recent Claude Jailbreak Incidents — rgblong · 2026-08-01