Liquid AI Releases Two Long-Context Encoders for CPU Workloads
The open-weight 230M and 350M models target classification, routing and extraction across inputs as long as 8,192 tokens.
Liquid AI released two open-weight LFM2.5 encoder models on July 28, extending its compact model family into long-context text understanding. The 230-million-parameter and 350-million-parameter models support inputs of as many as 8,192 tokens and are aimed at tasks such as classification, routing, extraction and scoring.
What changed
The company converted its existing LFM2.5 decoder backbones into bidirectional encoders. Its release notes describe changes to the attention mask and convolution layers, followed by masked-language training in two stages: a shorter 1,024-token phase for general language competence and an 8,192-token adaptation phase using a broader data mix.
Liquid AI also published results from 17 tasks drawn from GLUE, SuperGLUE and multilingual classification suites. The company says its scores are averaged across five held-out seeds. These are vendor-run measurements, so they establish the company’s evaluation record rather than an independent performance ranking.
Why it matters
Encoders are useful when an application needs to classify or score text without generating a response. Smaller models that can run on existing CPUs may lower the infrastructure barrier for high-volume filters, intent routers and document workflows, especially where sending text to an external generative service is undesirable.
Liquid AI reports its 230M model took about 28 seconds for one 8,192-token forward pass in its CPU test, compared with more than 90 seconds for ModernBERT-base. That is the basis for the company’s roughly 3.7-times speed claim. The post does not identify the processor used in the surrounding benchmark prose, so the timing should not be treated as portable across laptops and servers without independent testing.
Availability and limits
Both models are available as open weights on Hugging Face. Liquid AI also provides CPU-only demonstrations for prompt routing, policy linting, spell checking, personal-information detection and masked-diffusion text generation. Fine-tuning is still required for downstream task outputs, and the published benchmarks do not replace testing on a reader’s own data and hardware.
Status
Confirmed. The model files and first-party release post are public. Internal confidence is medium because availability is directly verifiable but performance and efficiency claims come from Liquid AI’s own evaluation.
Sources
Update note: Last reviewed 2026-08-02. We will revise this post if Liquid AI publishes hardware details or independent evaluations materially change the comparison.
Sources
- Liquid AI team article on Hugging Face — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.