tessella-moe-bench ⚡

Synthetic 32L × 64E Qwen2-MoE model with random weights for Tessella benchmarking.

Same structure as Qwen1.5-MoE-A2.7B but with hidden_size=512 to fit in 4GB VRAM. Do not use for inference — weights are random, output is noise.

	Value
Layers	32
Experts	64 (2 active/token)
Hidden size	512
Size	~1.6GB F16

# worker.toml
default_model = "victorespada/tessella-moe-bench"