poisonxa/PXA-Fusion4-35B-GGUF
point at the current engine repo and Discord
Model card: PXQ2 single-16GB tier published (measured fit on a 16 GB card)
Add PXA-Fusion4-35B-PXQ2.gguf: 10.5 GB, MTP-headed, fits one 16 GB Tesla P100/V100; chat template with the 2026-07-24 thinking default
card: publish PXQ4 only - drop withdrawn tiers and vision claims
Add PXA-Fusion4-35B-PXQ4 (PXQ4 expert bank, MTP head)
Remove superseded quantizations
docs: correct agentic flags - reasoning-budget 0 measured inert, add penalties + hard token cap
docs: chat-template thinking-polarity fix notice + agentic launch guidance
fix: chat template enable_thinking polarity (thinking now opt-in, matching upstream Qwen)
fix: chat template enable_thinking polarity (thinking now opt-in, matching upstream Qwen)
fix: chat template enable_thinking polarity (thinking now opt-in, matching upstream Qwen)
fix: chat template enable_thinking polarity (thinking now opt-in, matching upstream Qwen)
fix: chat template enable_thinking polarity (thinking now opt-in, matching upstream Qwen)
model card: drop benchmark/eval numbers; keep pitch + quants + run config
model card: keep run config only; remove internal test-methodology section
docs: exact serving params + gauntlet/SWE test methodology (fixes looping reports)
Add PXA-Fusion4-35B-PXQ6.gguf
Add PXA-Fusion4-35B-PXQ4.gguf
Add PXA-Fusion4-35B-PXQU16.gguf
Delete _uptest2.bin with huggingface_hub
Upload _uptest2.bin with huggingface_hub
Delete _uptest.bin with huggingface_hub
Upload _uptest.bin with huggingface_hub
card: availability note (PXQU12/PXQ2 live, rest uploading)
card: real Fusion4 card (155/179 gauntlet, 39.3% SWE, quant+speed tables, discord/github)
Add PXA-Fusion4-35B-PXQ2.gguf
Add PXA-Fusion4-35B-PXQU12.gguf
Add mmproj-fusion4-f16.gguf
Add banner.png
Remove Fusion2 artifacts ahead of Fusion4 refresh
fix: strip internal codename metadata (general.name/finetune/imatrix provenance) from PXQU16
fix: strip internal codename metadata (general.name/finetune/imatrix provenance) from PXQU12
fix: strip internal codename metadata (general.name/finetune/imatrix provenance) from PXQ4
fix: strip internal codename metadata (general.name/finetune/imatrix provenance) from PXQ3
fix: strip internal codename metadata (general.name/finetune/imatrix provenance) from PXQ2
docs: clean model card (standardize Qwen3.5-35B-A3B base naming, add org/license header)
Upload PXA-Fusion2-35B-PXQU12.gguf with huggingface_hub
Upload PXA-Fusion2-35B-PXQU16.gguf with huggingface_hub
Upload PXA-Fusion2-35B-PXQ2.gguf with huggingface_hub
Upload PXA-Fusion2-35B-PXQ3.gguf with huggingface_hub
Upload PXA-Fusion2-35B-PXQ4.gguf with huggingface_hub
Upload README.md with huggingface_hub
Clean full model card
Full model card (replaces beta placeholder)
Add vision projector
Beta model card
Clear old GGUFs; rebuilding with upgraded model + new quants
README: add visual PXQ vs IQ_K benchmark scorecard
refresh model card: PXQU tiers + head-to-head
rename universal tiers -> PXQU-16/PXQU-12