Liquid AI Releases LFM2.5: Five 1B Models, One Architecture

What It Is

Liquid AI has released LFM2.5, a collection of five specialized models that share a common foundation: a ~1 billion parameter architecture trained on 28 trillion tokens. The collection includes a general-purpose instruction model, a Japanese-optimized chat variant, a vision-language model capable of processing images, an audio-native model for speech tasks, and a base checkpoint for developers who want to fine-tune their own variants.

This release builds on Liquid AI’s earlier LFM2 architecture but incorporates significant improvements in instruction-following capabilities through expanded reinforcement learning during post-training. The models are designed specifically for on-device deployment, prioritizing low latency and offline functionality over cloud-dependent inference. All five models are available as open-weight releases at https://huggingface.co/collections/LiquidAI/lfm25, meaning developers can download and deploy them without ongoing API costs.

Why It Matters

The LFM2.5 collection addresses a growing demand for capable models that run locally rather than through cloud APIs. For mobile app developers, edge computing projects, and privacy-sensitive applications, this release provides production-ready options that eliminate network latency and reduce operational costs. The 1B parameter size hits a sweet spot - small enough to run on consumer hardware but large enough to handle real-world tasks effectively.

The multimodal approach is particularly noteworthy. Rather than forcing developers to stitch together separate models for text, vision, and audio tasks, Liquid AI provides specialized variants that share architectural DNA. This consistency simplifies deployment pipelines and makes it easier to maintain applications that need multiple modalities.

The Japanese-optimized variant signals attention to non-English markets, an area where many model providers still lag. Teams building products for Japanese users gain a model specifically tuned for that language rather than relying on general multilingual models that often underperform on specific languages.

The training scale - 28 trillion tokens compared to the previous 10 trillion - suggests Liquid AI invested heavily in data quality and quantity. This expanded training corpus likely contributes to the improved instruction-following behavior, making the models more reliable for production use cases where unpredictable outputs create real problems.

Getting Started

Developers can access the models through the Hugging Face collection at https://huggingface.co/collections/LiquidAI/lfm25. For basic inference with the standard instruct model:

model = AutoModelForCausalLM.from_pretrained("LiquidAI/lfm25-instruct")
tokenizer = AutoTokenizer.from_pretrained("LiquidAI/lfm25-instruct")

prompt = "Explain how transformers work in machine learning"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=200)
print(tokenizer.decode(outputs[0]))

For vision tasks, the vision-language variant accepts image inputs alongside text prompts. The audio-native model processes speech directly without requiring separate transcription steps, which reduces latency for voice-based applications.

Teams interested in custom fine-tuning should start with the base checkpoint, which provides the pre-trained foundation without the instruction-tuning layer. This approach works well for domain-specific applications where general instruction-following isn’t the primary goal.

Context

The 1B parameter range has become increasingly competitive. Microsoft’s Phi series, Google’s Gemma models, and various Qwen variants all target similar deployment scenarios. LFM2.5 differentiates itself through the multimodal variants and the explicit focus on on-device performance rather than benchmark scores.

Compared to larger models like Llama 3.1 8B or Mistral 7B, the LFM2.5 models trade some capability for dramatically lower resource requirements. Applications that previously needed cloud inference or expensive hardware can now run on mid-range devices.

The open-weight licensing removes barriers for commercial deployment, though developers should verify the specific license terms for their use case. The lack of API costs makes these models particularly attractive for high-volume applications where per-token pricing would become prohibitive.

One limitation: 1B models still struggle with complex reasoning tasks and specialized knowledge domains. Teams building applications that require deep expertise or multi-step logical inference may need to combine these models with retrieval systems or fall back to larger models for specific tasks.