TL;DR
Liquid AI has released 230-million- and 350-million-parameter LFM2.5 encoder models with an 8,192-token context window. The company says the smaller model is about 3.7 times faster than ModernBERT-base on long inputs using a CPU, but independent testing is not yet available.
Liquid AI has released two general-purpose language encoders, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, for classification, extraction, routing and other text-processing jobs. Both support an 8,192-token context window, while the company says their architecture delivers faster long-input inference on ordinary CPUs than larger competing models.
The models are available through Hugging Face and were derived from Liquid AI’s LFM2.5-230M and LFM2.5-350M decoder backbones. Liquid AI converted those causal models into bidirectional encoders by changing their attention masks, applying symmetric padding to short convolutions and training them with 30% of input tokens masked.
Training was conducted in two stages, according to the company. The models first learned through a masked-language objective using 1,024-token sequences from a large web corpus. Liquid AI then extended them to 8,192 tokens using a broader data mix intended to improve factual, legal and multilingual performance. These are general-purpose encoders, rather than systems restricted to document retrieval.
Liquid AI evaluated fully fine-tuned versions on 17 tasks drawn from GLUE, SuperGLUE and multilingual classification benchmarks. The company reported averages across five held-out seeds and said the 350M model ranked fourth among 14 tested models. It also said the 230M model outscored ModernBERT-base and every EuroBERT model in the comparison. The evaluation framework and raw results were released for inspection, but the supplied material does not include independent replication.
LFM2.5 Encoders for Fast Long-Context CPU Inference
Liquid AI’s new 230M and 350M bidirectional encoders target classification, extraction, routing and document understanding—supporting up to 8,192 tokens without requiring a dedicated accelerator.
230M model versus ModernBERT-base on an 8,192-token forward pass.
Built for contracts, transcripts, support threads and other long inputs.
Throughput-focused and accuracy-focused deployment options.
One encoder family, two deployment priorities
Both models originate from LFM2.5 decoder backbones converted into bidirectional encoders. The practical choice is primarily between throughput and task accuracy.
LFM2.5-Encoder-230M
Liquid AI’s recommended option for constrained hardware, recurring document workloads and higher CPU throughput.
LFM2.5-Encoder-350M
The larger model is positioned for teams willing to spend more compute in pursuit of stronger task-specific accuracy.
The CPU speed claim in perspective
The headline result comes from Liquid AI’s comparison at the maximum 8,192-token context length. It is promising—but remains a company-reported measurement.
Forward-pass time at 8,192 tokens
Lower is better · approximate time reported by Liquid AI
From causal decoder to bidirectional encoder
Liquid AI adapted existing LFM2.5 backbones, then used masked-language training to produce models suited to understanding and labeling text.
LFM2.5 decoder backbone
The 230M and 350M causal models provide the initial architecture.
Bidirectional attention
Attention masks were changed so tokens can use context from both directions.
Symmetric padding
Short convolutions received symmetric padding for encoder-style processing.
30% masked tokens
The models learned to reconstruct masked inputs rather than predict only forward.
Foundation training
Masked-language objective using 1,024-token sequences from a large web corpus.
Long-context extension
Training expanded to 8,192 tokens with a broader factual, legal and multilingual mix.
What is known—and what still needs proof
The release includes an evaluation framework and raw results, improving inspectability. Independent reproduction is still the missing layer.
| Claim or capability | LFM2.5 encoders | Evidence supplied | Current confidence |
|---|---|---|---|
| 8,192-token input support | Both models | Model release and documentation | Documented |
| 3.7× CPU speed advantage | 230M at maximum context | Liquid AI benchmark | Vendor claim |
| Strong classification quality | 17-task fine-tuned evaluation | Five held-out seeds averaged | Needs replication |
| Independent benchmark confirmation | Not yet available | No third-party reproduction supplied | Missing |
| Production behavior under quantization | Not established | No complete deployment comparison | Unknown |
Inspectible evaluation
Liquid AI released its evaluation framework and raw results for developer review.
Hardware variance
Different CPUs, runtimes, batch sizes and numerical formats may shift latency considerably.
Production economics
Memory use, sustained throughput, fine-tuning cost and accuracy on specialized documents need measurement.
Designed for understanding, not open-ended generation
Encoders can be more resource-efficient than generative models when the task is to score, classify, route or label existing text.
Support-thread routing
Assign long conversations to teams, intents, urgency levels or automated workflows.
Policy and contract checks
Score long documents for policy violations, clauses, obligations or review priority.
Personal-information detection
Identify token-level entities; Liquid AI demonstrated 40 PII categories across 16 languages.
Filtering and moderation
Attach task-specific heads to classify content without producing a generated response.
Prompt and intent routing
Direct requests to specialized models, agents or services using CPU-only inference.
Token labeling and correction
Support spelling correction, entity extraction and multilingual sequence analysis.
From release claim to deployment decision
The bottom line
LFM2.5-Encoders are a credible new option for long-document CPU processing. Their compact sizes, 8,192-token context and general-purpose design make them relevant to classification, extraction and routing pipelines. The reported 3.7× speedup is the central attraction, but it should remain a benchmark hypothesis until reproduced across independent hardware and production software stacks.
Source: Liquid AI release announcement and Hugging Face materials · Vetted by the thorstenmeyerai.com teamCPU Speed for Document Workloads
The main attraction is the reported performance on long CPU workloads. At an input length of 8,192 tokens, Liquid AI said ModernBERT-base required more than 90 seconds for one forward pass, compared with about 28 seconds for LFM2.5-Encoder-230M. That represents the company’s reported 3.7-fold speed advantage.
If reproduced under production conditions, those results could make document-scale classification practical without dedicated accelerators. Organizations could run contract scanning, support-thread classification, policy checks and personal-information detection on existing CPU systems. Fine-tuned encoders may also use fewer computing resources than generative large language models when the job involves scoring or labeling text rather than producing an open-ended response.
The company reported a narrower advantage on graphics processors. ModernBERT-base led below roughly 1,000 tokens on an Apple GPU, while the LFM2.5 encoders moved ahead at around 2,000 tokens. The announcement is focused on workloads where input length and recurring inference cost matter more than short-prompt latency.
CPU long document processing AI encoder
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Retrievers to General Encoders
The release follows Liquid AI’s LFM2.5-Retrievers, which were developed for multilingual search. The new models belong to the same family but use masked-language pretraining so developers can adapt them for sequence classification, token-level labeling and retrieval. Liquid AI cited intent routing, safety filtering and text classification as intended production uses.
The company also published CPU-only demonstrations covering zero-shot prompt routing, policy linting, spelling correction and detection of 40 categories of personal information across 16 languages. A separate demonstration uses iterative unmasking for text generation, though generation is not the primary role described for the encoders.
“Today, we release two new encoder models on Hugging Face: LFM2.5-Encoder-230M and LFM2.5-Encoder-350M.”
— Liquid AI, in its release announcement
long context language model CPU inference
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Testing Is Still Missing
It is not yet clear how the models will perform across different CPU architectures, batch sizes, numerical formats and production software stacks. The supplied material does not give all hardware and configuration details needed to determine whether users should expect the reported 28-second long-context result on their own machines.
The benchmark scores are also company-reported results from fully fine-tuned models. Independent comparisons could clarify memory consumption, fine-tuning cost, performance on specialized documents and accuracy against a broader set of recent encoders. The announcement does not establish how much quality may change under quantization or other deployment settings.
Fault Detection and Diagnosis in Industrial Systems
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Developers Begin Real-World Evaluation
Developers can now load the models through the Transformers library, run masked-token prediction or attach task-specific heads for fine-tuning. Liquid AI recommends the 350M model when accuracy has priority and the 230M model for tighter hardware or higher throughput. The next evidence will come from independent benchmark reproduction and tests on sustained document-processing workloads.
The Mathematics of Large Language Models: Machine Learning Theory Made Readable: LLMs, Transformers, Diffusion, Neural Networks, Optimization, and Generative AI
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What are LFM2.5-Encoder-230M and LFM2.5-Encoder-350M?
They are bidirectional language encoders from Liquid AI with 230 million and 350 million parameters. They are built for understanding tasks such as classification, routing, extraction, token labeling and search.
How much text can the models process?
Both models support inputs of up to 8,192 tokens. That capacity is intended for long contracts, transcripts, support conversations and other document-scale inputs.
Are they really 3.7 times faster than ModernBERT-base?
Liquid AI reports that LFM2.5-Encoder-230M was about 3.7 times faster at 8,192 tokens in its CPU comparison. The supplied source does not provide independent confirmation, so the result should be treated as a vendor benchmark claim.
Which LFM2.5 encoder should developers choose?
Liquid AI positions LFM2.5-Encoder-350M as the accuracy-focused option and LFM2.5-Encoder-230M as the choice for higher throughput or more limited hardware. Actual selection will depend on task-specific testing.
Source: Hugging Face
