Qwen 3.5 is a new generation of large language models designed for high-quality reasoning and multimodal tasks.
FlashHead is Embedl’s drop-in replacement for the output (lm) head in language models. It reduces the cost of the final layer without retraining, improving inference speed while preserving model quality.
This post shows how the combination performs in practice, and why it works.
Results
We applied FlashHead across the Qwen 3.5 family (0.8B → 27B) and validated accuracy on lm-eval benchmarks as well as on-device performance across multiple devices.
*Screenshots from https://github.com/embedl/Edge-Inference-Benchmarks
What actually changes
- Up to 1.4× faster inference
- Largest gains on smaller and mid-size models (0.8B–9B)
- Diminishing but still meaningful gains at larger sizes (≥27B)
- No measurable regression in quality across evals
Where the speedup comes from
On edge hardware (Jetson Nano Super, Jetson AGX Orin, Jetson AGX Thor), the LM head is often a disproportionate bottleneck:
- Large vocabulary projection
- Memory-bound operations
- Poor hardware utilization
FlashHead targets exactly this.
Why this matters for Qwen 3.5
Qwen 3.5 already pushes efficiency hard:
- Small models punching above their size (e.g. 9B competing with much larger models on reasoning benchmarks)
- Strong multimodal capability in a relatively compact footprint
That means:
- You’re more likely to be bottlenecked on inference overheads (like the LM head)
- Optimizing that layer gives immediate, visible gains
Benchmarks and raw results
📊 https://github.com/embedl/Edge-Inference-Benchmarks
Prebuilt models
🤗 https://huggingface.co/collections/embedl/qwen35
Summary
- FlashHead is a drop-in latency optimization for LLMs
- Qwen 3.5 + FlashHead delivers faster inference with unchanged quality
If you're running models on edge or latency-constrained setups, this is one of the simplest wins available.
Try it out
FlashHead models for Qwen 3.5 are available now:
🤗 https://huggingface.co/collections/embedl/qwen35
Install FlashHead:
pip install flash-head
Code and integration:
https://github.com/embedl/flash-head