Post
53
---
🚀 Gemma-4-A4B 98e v7-coder cohort — loop-fixed re-release. Two 20.8B MoE coders (4B-active), fresh-map prunes of Gemma 4 26B-A4B, 30/128 experts dropped per layer. The headline isn't a benchmark: the agentic loop is
gone at the weights, not papered over by the sampler.
🔧 How: at prune time we force-keep the 46 agentic_eog experts a loop-protection signal flags as load-bearing for clean multi-turn termination (+ shared-FFN α=1.2). Result: 0 loops across 48 seeds on every published
tier.
📊 Q6_K · llama.cpp · greedy · same host (from summary.json):
⚖️ v7-coder (fkbroad code3/lcb2) — balanced coder: LCB-med-55 98.18, HumanEval 98.17, HE+ 92.07, AIME 80.0, MATH-500 95.0, GSM8K 91, IFEval 92, MultiPL-E 89.7, ARC 92.2.
⚡ v7-coderx (code4/lcb3) — code-maximal: all-hard LCB-77 85.71 (cohort-best; 128e 79.22, v7-coder 84.42), HE+ 93.29, GSM8K 93, MATH-500 95.0, AIME 76.67. Whole budget on code.
🎯 Both land near GPQA ~51 — graduate science is the budget axis, neither is a science model. Pick v7-coder for the broad LCB-medium + HumanEval lead; v7-coderx for the all-hard slice and HE+.
🧪 The harness we used to prove the fix is now an omk tool: agentic-loop-harness replays a frozen agentic conversation across a sampler×seed matrix and reports a fail-rate per chat-template, so you can isolate a loop
to one variable. Model-agnostic — any OpenAI-compatible server. The version we shared with Google: google/gemma-4-12B-it#41
📦 Each ships bf16 · GGUF (+ CD-* + imatrix + mmproj vision) · NVFP4A16 (~13 GB) · Ollama.
🔗 ManniX-ITA/gemma-4-A4B-98e-v7-coder-it (+ -it-GGUF, -NVFP4A16) · https://ollama.com/mannix/gemma4-98e-v7-coder
🔗 ManniX-ITA/gemma-4-A4B-98e-v7-coderx-it (+ -it-GGUF, -NVFP4A16) · https://ollama.com/mannix/gemma4-98e-v7-coderx
🔧 https://github.com/mann1x/omnimergekit/tree/main/tools/agentic-loop-harness
🚀 Gemma-4-A4B 98e v7-coder cohort — loop-fixed re-release. Two 20.8B MoE coders (4B-active), fresh-map prunes of Gemma 4 26B-A4B, 30/128 experts dropped per layer. The headline isn't a benchmark: the agentic loop is
gone at the weights, not papered over by the sampler.
🔧 How: at prune time we force-keep the 46 agentic_eog experts a loop-protection signal flags as load-bearing for clean multi-turn termination (+ shared-FFN α=1.2). Result: 0 loops across 48 seeds on every published
tier.
📊 Q6_K · llama.cpp · greedy · same host (from summary.json):
⚖️ v7-coder (fkbroad code3/lcb2) — balanced coder: LCB-med-55 98.18, HumanEval 98.17, HE+ 92.07, AIME 80.0, MATH-500 95.0, GSM8K 91, IFEval 92, MultiPL-E 89.7, ARC 92.2.
⚡ v7-coderx (code4/lcb3) — code-maximal: all-hard LCB-77 85.71 (cohort-best; 128e 79.22, v7-coder 84.42), HE+ 93.29, GSM8K 93, MATH-500 95.0, AIME 76.67. Whole budget on code.
🎯 Both land near GPQA ~51 — graduate science is the budget axis, neither is a science model. Pick v7-coder for the broad LCB-medium + HumanEval lead; v7-coderx for the all-hard slice and HE+.
🧪 The harness we used to prove the fix is now an omk tool: agentic-loop-harness replays a frozen agentic conversation across a sampler×seed matrix and reports a fail-rate per chat-template, so you can isolate a loop
to one variable. Model-agnostic — any OpenAI-compatible server. The version we shared with Google: google/gemma-4-12B-it#41
📦 Each ships bf16 · GGUF (+ CD-* + imatrix + mmproj vision) · NVFP4A16 (~13 GB) · Ollama.
🔗 ManniX-ITA/gemma-4-A4B-98e-v7-coder-it (+ -it-GGUF, -NVFP4A16) · https://ollama.com/mannix/gemma4-98e-v7-coder
🔗 ManniX-ITA/gemma-4-A4B-98e-v7-coderx-it (+ -it-GGUF, -NVFP4A16) · https://ollama.com/mannix/gemma4-98e-v7-coderx
🔧 https://github.com/mann1x/omnimergekit/tree/main/tools/agentic-loop-harness
