Efficient Drop-In Replacement for the Classification Head in Language Model Inference. https://github.com/embedl/flash-head
AI & ML interests
None defined yet.
Recent Activity
nvidia/Cosmos-Reason2 multi-modal reasoning models optimized by Embedl.
-
embedl/Cosmos-Reason2-2B-W4A16-Edge2
Image-Text-to-Text • 2B • Updated • 227 • 13 -
embedl/Cosmos-Reason2-2B-W4A16-Edge2-FlashHead
Image-Text-to-Text • 2B • Updated • 1.28k • 9 -
embedl/Cosmos-Reason2-2B-NVFP4A16
Image-Text-to-Text • 2B • Updated • 1 -
embedl/Cosmos-Reason2-2B-W4A16
Image-Text-to-Text • 2B • Updated • 1.4k • 11
Ultra-efficient model variants optimized for Jetson Orin Nano. Designed for constrained edge environments requiring low memory footprint.
-
embedl/Cosmos-Reason2-2B-W4A16-Edge2
Image-Text-to-Text • 2B • Updated • 227 • 13 -
embedl/Cosmos-Reason2-2B-W4A16
Image-Text-to-Text • 2B • Updated • 1.4k • 11 -
embedl/Cosmos-Reason2-2B-W4A16-Edge2-FlashHead
Image-Text-to-Text • 2B • Updated • 1.28k • 9 -
Edge Inference Benchmarks
🚀7On-Device benchmarks across devices and models.
Models validated and performance-optimized for NVIDIA Jetson AGX Thor. Tailored for high-performance edge AI workloads.
-
embedl/Cosmos-Reason2-2B-NVFP4A16
Image-Text-to-Text • 2B • Updated • 1 -
embedl/Cosmos-Reason2-2B-W4A16-Edge2
Image-Text-to-Text • 2B • Updated • 227 • 13 -
embedl/Cosmos-Reason2-2B-W4A16
Image-Text-to-Text • 2B • Updated • 1.4k • 11 -
embedl/Cosmos-Reason2-2B-W4A16-Edge2-FlashHead
Image-Text-to-Text • 2B • Updated • 1.28k • 9
Quantization strategy where most weights are converted to INT4, activations remain in FP16, and sensitive layers are preserved in FP16.
Qwen/Qwen3.5 variants optimized by embedl.
-
embedl/Qwen3.5-0.8B-FlashHead
Image-Text-to-Text • 0.9B • Updated • 25 • 1 -
embedl/Qwen3.5-2B-FlashHead
Image-Text-to-Text • 2B • Updated • 2 -
embedl/Qwen3.5-4B-FlashHead
Image-Text-to-Text • 5B • Updated • 9 • 2 -
embedl/Qwen3.5-9B-FlashHead
Image-Text-to-Text • 10B • Updated • 2
Models optimized and bench-marked for NVIDIA Jetson AGX Orin. Memory-efficient and latency-optimized variants designed for real-time edge inference.
-
embedl/Cosmos-Reason2-2B-W4A16-Edge2
Image-Text-to-Text • 2B • Updated • 227 • 13 -
embedl/Cosmos-Reason2-2B-W4A16
Image-Text-to-Text • 2B • Updated • 1.4k • 11 -
embedl/Cosmos-Reason2-2B-W4A16-Edge2-FlashHead
Image-Text-to-Text • 2B • Updated • 1.28k • 9 -
embedl/Qwen3.5-0.8B-FlashHead
Image-Text-to-Text • 0.9B • Updated • 25 • 1
Efficient Drop-In Replacement for the Classification Head in Language Model Inference. https://github.com/embedl/flash-head
Quantization strategy where most weights are converted to INT4, activations remain in FP16, and sensitive layers are preserved in FP16.
nvidia/Cosmos-Reason2 multi-modal reasoning models optimized by Embedl.
-
embedl/Cosmos-Reason2-2B-W4A16-Edge2
Image-Text-to-Text • 2B • Updated • 227 • 13 -
embedl/Cosmos-Reason2-2B-W4A16-Edge2-FlashHead
Image-Text-to-Text • 2B • Updated • 1.28k • 9 -
embedl/Cosmos-Reason2-2B-NVFP4A16
Image-Text-to-Text • 2B • Updated • 1 -
embedl/Cosmos-Reason2-2B-W4A16
Image-Text-to-Text • 2B • Updated • 1.4k • 11
Qwen/Qwen3.5 variants optimized by embedl.
-
embedl/Qwen3.5-0.8B-FlashHead
Image-Text-to-Text • 0.9B • Updated • 25 • 1 -
embedl/Qwen3.5-2B-FlashHead
Image-Text-to-Text • 2B • Updated • 2 -
embedl/Qwen3.5-4B-FlashHead
Image-Text-to-Text • 5B • Updated • 9 • 2 -
embedl/Qwen3.5-9B-FlashHead
Image-Text-to-Text • 10B • Updated • 2
Ultra-efficient model variants optimized for Jetson Orin Nano. Designed for constrained edge environments requiring low memory footprint.
-
embedl/Cosmos-Reason2-2B-W4A16-Edge2
Image-Text-to-Text • 2B • Updated • 227 • 13 -
embedl/Cosmos-Reason2-2B-W4A16
Image-Text-to-Text • 2B • Updated • 1.4k • 11 -
embedl/Cosmos-Reason2-2B-W4A16-Edge2-FlashHead
Image-Text-to-Text • 2B • Updated • 1.28k • 9 -
Edge Inference Benchmarks
🚀7On-Device benchmarks across devices and models.
Models optimized and bench-marked for NVIDIA Jetson AGX Orin. Memory-efficient and latency-optimized variants designed for real-time edge inference.
-
embedl/Cosmos-Reason2-2B-W4A16-Edge2
Image-Text-to-Text • 2B • Updated • 227 • 13 -
embedl/Cosmos-Reason2-2B-W4A16
Image-Text-to-Text • 2B • Updated • 1.4k • 11 -
embedl/Cosmos-Reason2-2B-W4A16-Edge2-FlashHead
Image-Text-to-Text • 2B • Updated • 1.28k • 9 -
embedl/Qwen3.5-0.8B-FlashHead
Image-Text-to-Text • 0.9B • Updated • 25 • 1
Models validated and performance-optimized for NVIDIA Jetson AGX Thor. Tailored for high-performance edge AI workloads.
-
embedl/Cosmos-Reason2-2B-NVFP4A16
Image-Text-to-Text • 2B • Updated • 1 -
embedl/Cosmos-Reason2-2B-W4A16-Edge2
Image-Text-to-Text • 2B • Updated • 227 • 13 -
embedl/Cosmos-Reason2-2B-W4A16
Image-Text-to-Text • 2B • Updated • 1.4k • 11 -
embedl/Cosmos-Reason2-2B-W4A16-Edge2-FlashHead
Image-Text-to-Text • 2B • Updated • 1.28k • 9