This 35B Model Builds Full Apps on a $1.52/hr GPU (Nex-N2-mini)

A canvas fractal animation, written by a 35B model where only 3B params fire per token — one prompt, zero edits. This is Nex-N2-mini (built on Qwen3.5), and I rented a single A100 on RunPod ($1.52/hr) to run a full app-generation gauntlet on it. Everything on screen is real terminal output — including the parts where it failed.

The gauntlet

Canvas fractal, a complete 68KB finance dashboard, a neon Breakout game, tool calling, a hallucination trap, and a hard algorithm tested against brute force. The honest results:

• Two generations thought themselves to death (105k+ chars of reasoning blew the context window)
• The game needed three repair rounds; the dashboard shipped one bug
• The twist: I pasted every error back in — it fixed all 5 of its own bugs from the raw error text alone

Measured numbers

• 101 tok/s on 1× A100 80GB (Q8_0, Ollama 0.30.7)
• 36.9GB GGUF downloaded in 2m16s (~290MB/s)
• Hard algo: 5/5 correct vs brute force
• Adaptive thinking: 68 chars of thought (trivia) vs 1,009 (logic puzzle)

Model → https://huggingface.co/nex-agi/Nex-N2-mini · GGUF → https://huggingface.co/hendrik289/Nex-N2-mini-Q8_0-GGUF · Get RunPod

Full gauntlet in the video above. Subscribe for more real local-LLM tests.