TwoTiny LLMs are Redefining Edge AI: Less than 1-Gigabytes

To learn more about Local AI topics, check out related posts in the Local AI Series 

Have questions, ideas to share, or just want to connect? I’d love to hear from you! Check out my About Page to learn more about me or connect with me.

Disclaimer: I create this content entirely on my own time, and the views expressed here are mine alone (not my employer’s). Because I love leveraging new tech, I use AI tools like Gemini, NotebookLM, Claude, Perplexity and others as a “digital team” to help research and polish these articles so I can share the best possible insights with you!

There is a quiet revolution happening right under our noses—or more accurately, right inside our pockets.

For years, the narrative around Artificial Intelligence was “bigger is better.” Tech giants raced to build models with hundreds of billions of parameters, requiring massive server farms and nuclear-levels of electricity to run. But in early 2026, the pendulum swung violently toward the edge. Developers began asking a different question: How much intelligence can we squeeze into a single gigabyte?

A couple of months ago I wrote how: Small Local Models: Why Tiny AI Is Having a Big Moment

Two recent releases have turned this question into a fascinating architectural war.

In one corner, we have Liquid AI’s LFM2.5-1.2B-Thinking (released January 20, 2026). In the other, Alibaba Cloud’s Qwen3.5-0.8B (released March 2, 2026). Both fit entirely within the memory footprint of a standard budget smartphone, yet they represent two completely opposite bets on what a “tiny model” should be.

Two Divergent Philosophies: Reasoning vs. Versatility

When you only have about 1 GB of memory to play with after quantization (the process of compressing model weights), every single parameter counts. The creators of these models chose entirely different ways to spend their limited parameter budget.

1. LFM2.5-1.2B-Thinking: The Lean Reasoning Specialist

Liquid AI didn’t just build a smaller version of a standard model; they changed the blueprint entirely. Using a hybrid architecture that mixes gated convolution blocks with traditional attention mechanisms, LFM2.5 is designed strictly for text-based intelligence.

It pours its entire 1.17 billion parameters into deep reasoning, math, tool use, and instruction-following. It doesn’t care about images or videos. It is designed to think fast and accurately on text-only tasks.

2. Qwen3.5-0.8B: The Multimodal Swiss Army Knife

Alibaba took the exact opposite approach. Using a highly efficient hybrid architecture called DeltaNet, they crammed a staggering amount of features into just 800 million parameters.

Qwen3.5 doesn’t just read text; it processes images and video natively. It also boasts a massive 262,144-token context window and support for 201 languages. It is an ultra-compact generalist built to handle the chaotic, messy data of the real world.

The Spec Sheet

Feature	LFM2.5-1.2B-Thinking	Qwen3.5-0.8B
Release Date	January 20, 2026	March 2, 2026
Parameters	1.17 Billion	800 Million
Input Modalities	Text Only	Text, Image, Video
Context Window	32,768 tokens	262,144 tokens
Language Support	8 Languages	201 Languages
Typical Local Size	~696 MB	~1.0 GB
License	LFM Open License v1.0	Apache 2.0

Performance in the Wild: Speed vs. Capability

On-device benchmarks reveal the true cost of these architectural choices.

Because LFM2.5 skips the complex vision processing layers, it operates at lightning speeds on low-power hardware. When run on a modern phone CPU (like the Samsung Galaxy S25 Ultra), it blazes through text generation at roughly 70 tokens per second while pulling less than 719 MB of RAM. This makes it an absolute dream for latency-sensitive local tasks like real-time offline voice assistants, local data extraction, and rapid device automation.

Qwen3.5 carries a bit more structural weight. Its vision stack and massive context capacity mean it requires a slightly heavier 1.0 GB local memory footprint. On constrained edge hardware like a Raspberry Pi 5, it naturally runs slower than the nimble LFM. However, the trade-off is unparalleled capability for its size: it can look at a live camera feed, process an entire document library in a single prompt, or seamlessly translate highly localized languages completely offline.

Which “Tiny” Model Deserves Your Gigabyte?

For developers and privacy advocates looking to deploy local AI, the choice isn’t about which model is “better”—it’s about defining your deployment bottlenecks.

• Choose LFM2.5-1.2B-Thinking if you are building application logic, offline agents, or voice-first tools where speed, math, and strict instruction-following are your top priorities. It gives you the highest “reasoning per watt.”
• Choose Qwen3.5-0.8B if you are building an app that interacts with the physical world through images or video, or if you need to ingest massive logs of text locally without running out of memory.

The fact that we are even having a competitive debate between a local reasoning engine and a local multimodal giant—both running comfortably on a phone processor—proves that local AI is no longer a compromised novelty. It’s a production-ready reality.

Good to try them!