While the industry keeps chasing trillion-parameter frontier models, Liquid AI just shipped two open-weight bidirectional encoders that top out at 350 million parameters. The LFM2.5-Encoder-230M and LFM2.5-Encoder-350M carry an 8,192-token context window and, crucially, are tuned to stay quick on a plain CPU. It’s a deliberately unfashionable release — and probably a smart one.

Why encoders, and why now

Most of the headline hype belongs to decoders — the generative, chat-style models that write essays and code. Encoders do the quieter work: they turn text into vectors for retrieval, classification, reranking, and semantic search. That’s the plumbing behind RAG pipelines, recommendation systems, and content moderation. A faster, cheaper encoder doesn’t produce a viral demo, but it lowers the cost of nearly every retrieval-augmented system built on top of it.

Bidirectional matters here. Unlike a decoder that reads left-to-right, a bidirectional encoder sees the whole sequence at once, which typically yields richer representations for understanding tasks. Liquid AI builds these on its LFM2 hybrid backbone — the same lineage behind its efficiency-focused work — and the payoff shows up in the benchmarks.

The numbers that matter

On a 17-task suite spanning GLUE, SuperGLUE, and multilingual evaluations, the 350M model ranks fourth of 14 models tested, trailing only larger competitors. Punching above your weight class is the whole pitch: near-top-tier quality at a fraction of the parameter count.

The efficiency claim is the more interesting one. The 230M model clears a full 8K-token forward pass on CPU in roughly 28 seconds. That’s not GPU-fast, but it doesn’t need to be. It means a developer can run long-context embedding on commodity hardware — a laptop, a cheap VM, an edge device — without provisioning accelerators. For teams processing documents in batch rather than serving real-time chat, 28 seconds for 8,000 tokens on a CPU is entirely workable, and it removes GPU rental from the cost equation.

The bigger read

This fits a pattern that’s been building all year: the quiet migration of useful AI off the GPU and onto whatever silicon is already sitting on a desk. Long context has mostly been marketed as a decoder feature — million-token windows for reasoning. Liquid AI is arguing it belongs in the encoder too, where 8K tokens covers most real documents you’d actually want to embed whole rather than chunk into fragments.

Open weights sharpen the appeal. Anyone can drop these into an existing pipeline, fine-tune them, or run them fully on-premises for privacy-sensitive data that can’t leave the building. The competition here — established sentence-transformer families and other small embedding models — is real, and independent replication of the CPU timings will be the true test.

But the strategic bet is clear: not everyone needs the biggest model. A lot of people just need a fast, capable one that runs where they already are.