Nartex's view on AI - Chinese models are actually cheaper and winning.
To give an example, the Deepseek V4 model is between 20 and 50 times cheaper than Anthropic and that is why in programming it has increased its market share from 3% to 49% in volume.
OpenAI's proprietary inference AI chip "Jalapeno"
I have run a test on Token/Second specifically running Deepseek V4 Flash in Deepseek Harness
Runinfrai claims 250 token/sec while CommandCodeAI have no claims about speed, yet I saw really similar results. RunInfra: 2 turns · 73 steps · TTFT avg 2.8s · 108 tok/s cache hit 90%
My minimax h3 skills are evolving.
This is low quality config run with Deepseek-v4-flash co-tenant + Hermes orchestrating ComfyUI through their OSS MCP server. Took 33-minutes to create 5 x 5s clips with persistent voice and character. Hermes sticking the videos together.
You have reached the end of the archive
All of deepseek