tiyuvta/Phonon-2-ONNX
Automatic Speech Recognition • Updated • 188 • 3
AI & ML interests
AI integration for non-tech companies. Tailored systems, personal guidance and private deployment options. Based in Israel, in Hebrew and English. The models here are research releases from our work on making AI faster, cheaper and more accurate.
Recent Activity
View all activity
Avifenesh
updated 17
models 3 days ago
tiyuvta/MiMo-V2.6-Flash-RL-NVFP4
Text Generation • 311B • Updated • 456 • 3
tiyuvta/DictaLM-3.0-24B-NVFP4-healed
Text Generation • 24B • Updated • 319 • 2
tiyuvta/DictaLM-3.0-24B-NVFP4
Text Generation • 24B • Updated • 304
tiyuvta/GLM-5.3-Flash-NVFP4-B200-hybrid
Text Generation • 184B • Updated • 661 • 1
tiyuvta/GLM-5.3-Flash-NVFP4-4o6
Text Generation • 321B • Updated • 379 • 1
tiyuvta/Hy3-NVFP4
Text Generation • 299B • Updated • 477 • 1
tiyuvta/Qwen3.8-Flash-Next-NVFP4
Image-Text-to-Text • 117B • Updated • 139 • 1
tiyuvta/GLM-5.3-Flash-NVFP4
Text Generation • 321B • Updated • 363 • 1
tiyuvta/Qwen3.8-27B-DFlash2-memra
Text Generation • 2B • Updated • 698 • 2
tiyuvta/Ornith-1.5-9B-NVFP4-MTP-GGUF
Text Generation • 32.8k • Updated • 808 • 4
tiyuvta/gemma-4-31B-it-assistant-MTP-GGUF
Text Generation • 0.5B • Updated • 442 • 1
tiyuvta/Qwen3.8-27B-DSpark-Agentic
Text Generation • 1B • Updated • 676 • 5
tiyuvta/Ornith-1.5-35B-A3B-NVFP4-MTP-GGUF
Text Generation • 67.1M • Updated • 12k • 6
tiyuvta/DeepSeek-V4-Flash-0731-NVFP4
Text Generation • 304B • Updated • 267
tiyuvta/Qwen3.8-27B-NVFP4-MTP-GGUF
Text Generation • 32.8k • Updated • 1.08k • 15
tiyuvta/memra-bench
32.8k • Updated • 1.67k • 3
Avifenesh
updated a
Space 13 days ago
Post
2,204
I added a small gate to the Gated DeltaNet layers of Qwen3.5 and trained it with GRPO under a length penalty.
On Qwen3.5-4B it reasons 48% shorter than LoRA on MATH-500, and anytime accuracy goes from 0.786 to 0.852. A LoRA with about the same parameter count barely moves.
The gate sits on beta, the per-token write strength. One matrix per layer, zero init, so it starts as a no-op.
It started as a negative eigenvalue experiment. The negative range turned out not to matter. The gate did.
Write-up: https://huggingface.co/blog/Avifenesh/beta-gate-qwen35
Adapters: Avifenesh/gate-qwen35-6accb63081800b2f2d583b80
Code: https://github.com/avifenesh/gdn-beta-gatehttps://huggingface.co/collections/Avifenesh/gate-qwen35
On Qwen3.5-4B it reasons 48% shorter than LoRA on MATH-500, and anytime accuracy goes from 0.786 to 0.852. A LoRA with about the same parameter count barely moves.
The gate sits on beta, the per-token write strength. One matrix per layer, zero init, so it starts as a no-op.
It started as a negative eigenvalue experiment. The negative range turned out not to matter. The gate did.
Write-up: https://huggingface.co/blog/Avifenesh/beta-gate-qwen35
Adapters: Avifenesh/gate-qwen35-6accb63081800b2f2d583b80
Code: https://github.com/avifenesh/gdn-beta-gatehttps://huggingface.co/collections/Avifenesh/gate-qwen35
Avifenesh
published a
model about 1 month ago