LFM2.5 1.2B vs Llama 3.2 1B

Two 1B-class models that both fine-tune on a free T4. Instruction following, memory footprint on real phones, context length, licensing, and which one to pick for on-device work.

Overview

LFM2.5 1.2B and Llama 3.2 1B are the two 1B-class bases in the Ertas catalogue as of August 2026. Both train on the entry T4 tier, both sit inside the Free plan, and both export to GGUF at well under a gigabyte. On paper they look interchangeable. In practice they are built for different jobs, and the benchmark that separates them is the one most on-device projects actually care about.

Llama 3.2 1B is a dense transformer with a 128K-token context window and the deepest ecosystem of any small open-weight model. If a fine-tuning recipe, a quantisation preset, or a mobile integration guide exists for a 1B model, it almost certainly exists for this one.

LFM2.5 1.2B is a hybrid. It interleaves double-gated short-range convolution blocks with grouped-query attention, 16 layers split 10 convolution to 6 attention. Convolution costs linear time in sequence length where attention costs quadratic, which is what lets a model this small stay cheap on longer prompts. Its context window is 32,768 tokens, a quarter of Llama's, and its vocabulary is 65,536 tokens, which is a large part of why its exports are small.

Feature Comparison

Feature	LFM2.5 1.2B	Llama 3.2 1B
Parameters	1.17B	1.0B
Architecture	Hybrid: convolution + grouped-query attention	Dense transformer
Context window	32,768 tokens	128,000 tokens
Vocabulary size	65,536	128,256
IFEval (instruction following)	86.23	52.37
Q4_K_M export size	731 MB	810 MB
Measured phone footprint	719 MB (Galaxy S25 Ultra)	Not published by Meta
License	LFM Open License v1.0	Llama Community License
GPU tier to fine-tune on Ertas	T4	T4
Trainable on the Ertas Free plan
Native tool-call tokens	Yes, Pythonic calls by default	Prompt-formatted
Knowledge cutoff	Mid-2024	December 2023
Ecosystem maturity	Young, released 2026	Deepest of any 1B model

Strengths

LFM2.5 1.2B

Instruction following is the headline gap. LFM2.5 1.2B scores 86.23 on IFEval against 52.37 for Llama 3.2 1B, measured by Liquid AI as of August 2026. For a model whose job is to emit a parseable JSON object or a well-formed tool call every time, that axis outweighs trivia recall.
Tool calling is first-class rather than prompt-engineered. LFM2.5 writes Pythonic function calls between dedicated tool-call tokens, and a system prompt switches it to JSON output.
Smaller exports at a larger parameter count. 731 MB at Q4_K_M against Llama's 810 MB, despite carrying 17 percent more parameters, because the 65,536-token vocabulary keeps the embedding tables small. Embedding tables dominate file size at this scale.
Published footprints on named hardware. Liquid measured 719 MB on a Galaxy S25 Ultra. Meta publishes no equivalent figure for Llama 3.2 1B, so sizing it means measuring it yourself.
bf16-native weights, so training stays at bf16 on Ertas rather than falling back to fp32 on Turing hardware.

Llama 3.2 1B

A 128K context window against 32,768. Four times the room, and the deciding factor for any task that puts a long document in the prompt rather than retrieving into it.
The deepest ecosystem of any 1B-class model. Existing fine-tunes, quantisation presets, mobile integration guides and community-validated recipes overwhelmingly target Llama first.
A license most legal teams have already reviewed. The Llama Community License carries a 700M monthly-active-user cap and attribution requirements, and it is familiar, where the LFM Open License v1.0 is Liquid's own document and needs a fresh read.
Broader general knowledge. Llama 3.2 1B was trained for general capability, where LFM2.5 explicitly trades world knowledge for instruction reliability.
Stronger on programming and mathematics, both of which Liquid names as tasks it does not recommend LFM2.5 for.

Which Should You Choose?

You are turning documents into structured JSON, or driving tools from a small modelLFM2.5 1.2B

This is the case LFM2.5 is built for. IFEval 86.23 against 52.37 is the difference between a model that reliably produces the shape you asked for and one you have to wrap in retry logic. The facts come from the document, so the weaker world knowledge costs nothing here.

Your prompt carries a long document, a full transcript, or an extended chat historyLlama 3.2 1B

128K against 32,768 tokens decides it. LFM2.5's hybrid architecture makes long prompts cheaper to process, but it cannot process a prompt that exceeds its window, and chunking to fit reintroduces the retrieval problem you were avoiding.

You are shipping a model inside a browser, a game, or a consumer app downloadLFM2.5 1.2B

731 MB against 810 MB is a modest gap on its own, but the published on-device measurements are the real advantage: you can size the deployment before writing any code. Chatty Valley ships an LFM2.5 1.2B fine-tune at 697 MB inside a Stardew Valley mod, running on the player's own CPU.

You need general knowledge, coding help, or mathematical reasoning from a 1B modelLlama 3.2 1B

Liquid explicitly does not recommend LFM2.5 for knowledge-intensive tasks or programming. If the model needs to know things rather than do things, Llama 3.2 1B is the better base at this size, and a larger base is better still.

You are inheriting an existing 1B pipeline, or your team has already cleared Llama's licenseLlama 3.2 1B

Ecosystem maturity and a pre-reviewed license are real costs to re-pay. Unless the instruction-following gap is load-bearing for your task, the migration rarely pays for itself.

Verdict

Pick LFM2.5 1.2B when the model's job is to follow instructions precisely inside a fixed, modest context: structured extraction, tool calling, on-device agents, character and game AI. The IFEval gap is large enough to change what you can ship without a retry wrapper, and the published hardware measurements make deployment sizing a lookup instead of an experiment.

Pick Llama 3.2 1B when context length decides the task, when you need general knowledge or code, or when an existing ecosystem and an already-reviewed license are worth more than the instruction-following gap. For many teams the honest answer is that Llama is the safer default and LFM2.5 is the better specialist.

Both fine-tune on a free T4 in Ertas Studio, so the cheapest way to resolve this is to train both on your own data and compare. At this size that is an afternoon, not a project.

How Ertas Fits In

Both models fine-tune on the T4 tier in Ertas Studio and both sit under the Free plan's parameter ceiling, so you can train either without paying anything. For sub-2B bases the training recipe that works is a higher learning rate and more epochs than a 7B model tolerates: 3e-4 to 5e-4, batch size 4, gradient accumulation 2, and 8 to 16 epochs. Small models need more passes before a behaviour sticks.

Exports come out as GGUF for llama.cpp, Ollama and LM Studio, or as safetensors LoRA adapters if you want to keep the adapter separate and merge it yourself. Two Ertas builds run on LFM2.5 in production and are worth reading before you pick a size: Chatty Valley is a 1.2B fine-tune quantised to 697 MB running inside a Stardew Valley mod, and Corporate Goblin is a 230M LoRA exported to q4 ONNX running entirely in the browser.

One licensing note that travels with your fine-tune either way. The LFM Open License v1.0 is Liquid AI's own document rather than Apache 2.0, and the Llama Community License carries its own conditions. A commercial product built on either base needs the license read rather than assumed.