Nate@nate_512

フォロー

DeepSeek V4.1 Flash NVFP4 locally

LOCAL INFERENCE JUST GOT A MASSIVE UPGRADE atomic_chat_hq did it again 💥 They’ve made DeepSeek V4.1 Flash much more practical to run on Blackwell by squeezing the 552B MoE model into NVFP4. The result? lower inference costs without giving up the model’s core capabilities. You still get a 1M context window and native visual understanding, while retaining 98% agreement on agentic dialogue. That makes this a super strong fit for local agentic workflows. Grab the weights on huggingface 🤗 (link in 🧵↓)

2d

Nate@nate_512

フォロー

DeepSeek V4.1 Flash NVFP4 locally

LOCAL INFERENCE JUST GOT A MASSIVE UPGRADE atomic_chat_hq did it again 💥 They’ve made DeepSeek V4.1 Flash much more practical to run on Blackwell by squeezing the 552B MoE model into NVFP4. The result? lower inference costs without giving up the model’s core capabilities. You still get a 1M context window and native visual understanding, while retaining 98% agreement on agentic dialogue. That makes this a super strong fit for local agentic workflows. Grab the weights on huggingface 🤗 (link in 🧵↓)

2d