Joakimpalm-Zen/Qwen3-30B-A3B-selective-attnQ8_0-expQ4_0-GGUF

Text Generation • 31B • Updated 14 days ago • 728

Note Selective precision (attention Q8_0, experts Q4_0), 17.99 GB. Passes the fidelity bar where the uniform Q4_K_M fails it. The lead artifact.

Joakimpalm-Zen/Qwen3-Coder-30B-A3B-Instruct-keep120-Q4_K_M-GGUF

Text Generation • 29B • Updated 14 days ago • 1.01k • 1

Note Expert-pruned (120 of 128 kept), Q4_K_M, 17.5 GB. Passes both the original and the current bar.

Joakimpalm-Zen/NVIDIA-Nemotron-Nano-9B-v2-Q8_0-GGUF

Text Generation • 9B • Updated 14 days ago • 385 • 1

Note Plain canonical Q8_0 of NVIDIA's base, the first Mamba-2 hybrid Runner serves. 5/6 greedy-identical vs llama.cpp.

Joakimpalm-Zen/gemma-4-E2B-it-Q4_0-GGUF

Text Generation • 5B • Updated 14 days ago • 358 • 2

Note The 2.63 GB quickstart file for an 8 GB machine. Fails the fidelity bar and the card says so.

Joakimpalm-Zen/Qwen3-4B-Runner-ToolUse-Q4_K_M

16.5M • Updated 14 days ago • 499

Note LoRA trained through the served Q4_K_M weights, byte-reproducible. Held-out exact tool-call 0.69 base, 1.00 adapted.

Joakimpalm-Zen/Qwen3-30B-A3B-derive-GGUF

31B • Updated 8 days ago • 72

Note Made by runner derive, which ships in Runner 1.11.0