MC7ever/lfm3-omni-dualhead
A single dual-head architecture for fast decisions and generation
1 min readupdated
The model combines two functions on a single shared Tri-hybrid LFM3-Omni core. The first part (System 1) handles instant classification and query routing without generating tokens, while the second part (System 2) performs full-scale text and speech generation. Thanks to memory sharing, the solution saves up to 50% of VRAM compared to running two separate models. This development is useful for developers of chatbots, traffic routing systems, and multimodal applications.
Key features
- Combining classification (System 1) and generation (System 2) on a shared backbone
- VRAM savings: requires about 4.5 GB in default mode or 1.8 GB in 4-bit
- No weight duplication due to the use of shared memory tensors
- Support for working with text, image, and audio input
What you can do
- Build a system for fast filtering and triage of client requests in milliseconds
- Run multimodal classification of images or audio clips
- Use the generative mode for full-fledged dialogue and text processing
Example
import torch
from transformers import AutoModel
model = AutoModel.from_pretrained(
"MC7ever/lfm3-omni-dualhead",
trust_remote_code=True,
device_map="auto"
)
result = model.system_one(state="Urgent: payment error", questions={"is_urgent": {"type": "boolean"}})
print(result)
Read more
DetailsBenchmarks: System 1 speed under 50 ms per decision.
Hardware requirements: ~4.5 GB VRAM for BF16, ~1.8 GB for 4-bit.
Base models: LiquidAI/d1-Omni-2B and LiquidAI/LFM3-Omni-2B.
License: Apache 2.0.










Comments