DeepSeek just dropped V4.1-Flash and despite using far fewer active parameters, it’s already getting dangerously close to much larger models.
552B MoE beats V4 Pro on performance, cost and speed only 8B active parameters for input, 16B for output native vision performance
FREE - The New DeepSeek for the next 48 Hours.
We will route all your requests for free next 48 hours, no limits. Try it out -
Your keys. Your rates. One endpoint.
Bring Your Own Keys is coming to ModelBeat. Connect Anthropic, OpenAI, Mistral, Groq, DeepSeek, and Vertex, then run them through one API. Pooled routing across your keys, or zero-custody passthrough at your negotiated rate. ModelBeat does
DeepSeek V4.1 Flash is live on RamenBench!
On Max thinking, it worked for 11m 37s, spent 155K tokens, and absolutely destroys V4 Flash and V4 Pro Its still not in my Top 5 models, but this was a huge jump for this test to say the least. The vision really helps it get a groove
You have reached the end of the archive
All of deepseek